InvokeAI/invokeai/frontend/web/src/services/api/schema.ts

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

12875 lines
357 KiB
TypeScript
Raw Normal View History

/**
* This file was auto-generated by openapi-typescript.
* Do not make direct changes to the file.
*/
2024-03-05 11:33:01 +00:00
export type paths = {
2024-03-05 11:33:01 +00:00
"/api/v1/utilities/dynamicprompts": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Parse Dynamicprompts
* @description Creates a batch process
*/
2024-03-05 11:33:01 +00:00
post: operations["parse_dynamicprompts"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v2/models/": {
2023-11-16 00:21:28 +00:00
/**
* List Model Records
* @description Get a list of models.
*/
2024-03-05 11:33:01 +00:00
get: operations["list_model_records"];
2023-11-16 00:21:28 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v2/models/get_by_attrs": {
2024-02-22 22:53:33 +00:00
/**
* Get Model Records By Attrs
* @description Gets a model by its attributes. The main use of this route is to provide backwards compatibility with the old
* model manager, which identified models by a combination of name, base and type.
*/
2024-03-05 11:33:01 +00:00
get: operations["get_model_records_by_attrs"];
2024-02-22 22:53:33 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v2/models/i/{key}": {
2023-11-16 00:21:28 +00:00
/**
* Get Model Record
* @description Get a model record
2023-11-16 00:21:28 +00:00
*/
2024-03-05 11:33:01 +00:00
get: operations["get_model_record"];
2023-11-16 00:21:28 +00:00
/**
* Delete Model
* @description Delete model record from database.
*
* The configuration record will be removed. The corresponding weights files will be
* deleted as well if they reside within the InvokeAI "models" directory.
2023-11-16 00:21:28 +00:00
*/
delete: operations["delete_model"];
2023-11-16 00:21:28 +00:00
/**
* Update Model Record
* @description Update a model's config.
2023-11-16 00:21:28 +00:00
*/
2024-03-05 11:33:01 +00:00
patch: operations["update_model_record"];
2023-11-16 00:21:28 +00:00
};
"/api/v2/models/scan_folder": {
/** Scan For Models */
2024-03-05 11:33:01 +00:00
get: operations["scan_for_models"];
};
"/api/v2/models/install": {
2023-11-16 00:21:28 +00:00
/**
* List Model Installs
* @description Return the list of model install jobs.
*
* Install jobs have a numeric `id`, a `status`, and other fields that provide information on
* the nature of the job and its progress. The `status` is one of:
*
* * "waiting" -- Job is waiting in the queue to run
* * "downloading" -- Model file(s) are downloading
* * "running" -- Model has downloaded and the model probing and registration process is running
* * "completed" -- Installation completed successfully
* * "error" -- An error occurred. Details will be in the "error_type" and "error" fields.
* * "cancelled" -- Job was cancelled before completion.
*
* Once completed, information about the model such as its size, base
* model and type can be retrieved from the `config_out` field. For multi-file models such as diffusers,
* information on individual files can be retrieved from `download_parts`.
*
* See the example and schema below for more information.
2023-11-16 00:21:28 +00:00
*/
get: operations["list_model_installs"];
/**
* Install Model
2024-02-15 11:15:21 +00:00
* @description Install a model using a string identifier.
*
* `source` can be any of the following.
*
* 1. A path on the local filesystem ('C:\users\fred\model.safetensors')
* 2. A Url pointing to a single downloadable model file
* 3. A HuggingFace repo_id with any of the following formats:
* - model/name
* - model/name:fp16:vae
* - model/name::vae -- use default precision
* - model/name:fp16:path/to/model.safetensors
* - model/name::path/to/model.safetensors
*
* `config` is an optional dict containing model configuration values that will override
* the ones that are probed automatically.
*
* `access_token` is an optional access token for use with Urls that require
* authentication.
*
* Models will be downloaded, probed, configured and installed in a
* series of background threads. The return object has `status` attribute
* that can be used to monitor progress.
*
* See the documentation for `import_model_record` for more information on
* interpreting the job information returned by this route.
*/
post: operations["install_model"];
/**
* Prune Model Install Jobs
* @description Prune all completed and errored jobs from the install job list.
*/
delete: operations["prune_model_install_jobs"];
};
"/api/v2/models/install/{id}": {
2024-01-22 22:28:24 +00:00
/**
* Get Model Install Job
2024-02-15 11:15:21 +00:00
* @description Return model install job corresponding to the given source. See the documentation for 'List Model Install Jobs'
* for information on the format of the return value.
2024-01-22 22:28:24 +00:00
*/
2024-03-05 11:33:01 +00:00
get: operations["get_model_install_job"];
2024-01-22 22:28:24 +00:00
/**
* Cancel Model Install Job
* @description Cancel the model install job(s) corresponding to the given job ID.
*/
2024-03-05 11:33:01 +00:00
delete: operations["cancel_model_install_job"];
2024-01-22 22:28:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v2/models/sync": {
/**
* Sync Models To Config
2024-01-22 22:28:24 +00:00
* @description Traverse the models and autoimport directories.
*
* Model files without a corresponding
* record in the database are added. Orphan records without a models file are deleted.
*/
2024-03-05 11:33:01 +00:00
patch: operations["sync_models_to_config"];
};
2024-03-05 11:33:01 +00:00
"/api/v2/models/convert/{key}": {
2024-02-15 11:15:21 +00:00
/**
* Convert Model
* @description Permanently convert a model into diffusers format, replacing the safetensors version.
* Note that during the conversion process the key and model hash will change.
* The return value is the model configuration for the converted model.
*/
2024-03-05 11:33:01 +00:00
put: operations["convert_model"];
2024-02-15 11:15:21 +00:00
};
"/api/v1/download_queue/": {
/**
* List Downloads
* @description Get a list of active and inactive jobs.
*/
2024-03-05 11:33:01 +00:00
get: operations["list_downloads"];
/**
* Prune Downloads
* @description Prune completed and errored jobs.
*/
2024-03-05 11:33:01 +00:00
patch: operations["prune_downloads"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/download_queue/i/": {
/**
* Download
* @description Download the source URL to the file or directory indicted in dest.
*/
2024-03-05 11:33:01 +00:00
post: operations["download"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/download_queue/i/{id}": {
/**
* Get Download Job
* @description Get a download job using its ID.
*/
2024-03-05 11:33:01 +00:00
get: operations["get_download_job"];
/**
* Cancel Download Job
* @description Cancel a download job using its ID.
*/
2024-03-05 11:33:01 +00:00
delete: operations["cancel_download_job"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/download_queue/i": {
/**
* Cancel All Download Jobs
* @description Cancel all download jobs.
*/
2024-03-05 11:33:01 +00:00
delete: operations["cancel_all_download_jobs"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/upload": {
/**
2023-08-22 05:15:00 +00:00
* Upload Image
* @description Uploads an image
*/
2024-03-05 11:33:01 +00:00
post: operations["upload_image"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/i/{image_name}": {
/**
2023-08-22 05:15:00 +00:00
* Get Image Dto
2023-07-12 15:15:09 +00:00
* @description Gets an image's DTO
*/
2024-03-05 11:33:01 +00:00
get: operations["get_image_dto"];
/**
2023-08-22 05:15:00 +00:00
* Delete Image
* @description Deletes an image
*/
2024-03-05 11:33:01 +00:00
delete: operations["delete_image"];
/**
2023-08-22 05:15:00 +00:00
* Update Image
* @description Updates an image
*/
2024-03-05 11:33:01 +00:00
patch: operations["update_image"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/intermediates": {
/**
* Get Intermediates Count
* @description Gets the count of intermediate images
*/
2024-03-05 11:33:01 +00:00
get: operations["get_intermediates_count"];
/**
2023-08-22 05:15:00 +00:00
* Clear Intermediates
* @description Clears all intermediates
*/
2024-03-05 11:33:01 +00:00
delete: operations["clear_intermediates"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/i/{image_name}/metadata": {
/**
2023-08-22 05:15:00 +00:00
* Get Image Metadata
* @description Gets an image's metadata
*/
2024-03-05 11:33:01 +00:00
get: operations["get_image_metadata"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/i/{image_name}/workflow": {
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** Get Image Workflow */
2024-03-05 11:33:01 +00:00
get: operations["get_image_workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/i/{image_name}/full": {
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Get Image Full
2023-07-12 15:15:09 +00:00
* @description Gets a full-resolution image file
*/
2024-03-05 11:33:01 +00:00
get: operations["get_image_full"];
/**
2023-08-22 05:15:00 +00:00
* Get Image Full
* @description Gets a full-resolution image file
*/
2024-03-05 11:33:01 +00:00
head: operations["get_image_full"];
2023-07-12 15:15:09 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/i/{image_name}/thumbnail": {
/**
2023-08-22 05:15:00 +00:00
* Get Image Thumbnail
* @description Gets a thumbnail image file
*/
2024-03-05 11:33:01 +00:00
get: operations["get_image_thumbnail"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/i/{image_name}/urls": {
/**
2023-08-22 05:15:00 +00:00
* Get Image Urls
* @description Gets an image and thumbnail URL
*/
2024-03-05 11:33:01 +00:00
get: operations["get_image_urls"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/": {
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/**
2023-08-22 05:15:00 +00:00
* List Image Dtos
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Gets a list of image DTOs
*/
2024-03-05 11:33:01 +00:00
get: operations["list_image_dtos"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/delete": {
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** Delete Images From List */
2024-03-05 11:33:01 +00:00
post: operations["delete_images_from_list"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/star": {
/** Star Images In List */
2024-03-05 11:33:01 +00:00
post: operations["star_images_in_list"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/unstar": {
/** Unstar Images In List */
2024-03-05 11:33:01 +00:00
post: operations["unstar_images_in_list"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/download": {
2023-10-12 12:16:29 +00:00
/** Download Images From List */
2024-03-05 11:33:01 +00:00
post: operations["download_images_from_list"];
2023-10-12 12:16:29 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/images/download/{bulk_download_item_name}": {
2024-02-20 11:15:14 +00:00
/**
* Get Bulk Download Item
* @description Gets a bulk download zip file
*/
2024-03-05 11:33:01 +00:00
get: operations["get_bulk_download_item"];
2024-02-20 11:15:14 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/boards/": {
/**
2023-08-22 05:15:00 +00:00
* List Boards
* @description Gets a list of boards
*/
2024-03-05 11:33:01 +00:00
get: operations["list_boards"];
/**
2023-08-22 05:15:00 +00:00
* Create Board
* @description Creates a board
*/
2024-03-05 11:33:01 +00:00
post: operations["create_board"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/boards/{board_id}": {
/**
2023-08-22 05:15:00 +00:00
* Get Board
* @description Gets a board
*/
2024-03-05 11:33:01 +00:00
get: operations["get_board"];
/**
2023-08-22 05:15:00 +00:00
* Delete Board
* @description Deletes a board
*/
2024-03-05 11:33:01 +00:00
delete: operations["delete_board"];
/**
2023-08-22 05:15:00 +00:00
* Update Board
* @description Updates a board
*/
2024-03-05 11:33:01 +00:00
patch: operations["update_board"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/boards/{board_id}/image_names": {
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
/**
2023-08-22 05:15:00 +00:00
* List All Board Image Names
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
* @description Gets a list of images for a board
*/
2024-03-05 11:33:01 +00:00
get: operations["list_all_board_image_names"];
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/board_images/": {
/**
2023-08-22 05:15:00 +00:00
* Add Image To Board
* @description Creates a board_image
*/
2024-03-05 11:33:01 +00:00
post: operations["add_image_to_board"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/**
2023-08-22 05:15:00 +00:00
* Remove Image From Board
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Removes an image from its board, if it had one
*/
2024-03-05 11:33:01 +00:00
delete: operations["remove_image_from_board"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/board_images/batch": {
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/**
2023-08-22 05:15:00 +00:00
* Add Images To Board
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Adds a list of images to a board
*/
2024-03-05 11:33:01 +00:00
post: operations["add_images_to_board"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/board_images/batch/delete": {
/**
2023-08-22 05:15:00 +00:00
* Remove Images From Board
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Removes a list of images from their board, if they had one
*/
2024-03-05 11:33:01 +00:00
post: operations["remove_images_from_board"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/version": {
2023-07-08 09:31:17 +00:00
/** Get Version */
2024-03-05 11:33:01 +00:00
get: operations["app_version"];
2023-07-12 15:56:40 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/app_deps": {
2023-12-06 12:33:54 +00:00
/** Get App Deps */
2024-03-05 11:33:01 +00:00
get: operations["get_app_deps"];
2023-12-06 12:33:54 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/config": {
2023-07-12 15:56:40 +00:00
/** Get Config */
2024-03-05 11:33:01 +00:00
get: operations["get_config"];
2023-07-08 09:31:17 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/logging": {
2023-07-20 15:45:54 +00:00
/**
2023-08-22 05:15:00 +00:00
* Get Log Level
2023-07-20 15:45:54 +00:00
* @description Returns the log level
*/
2024-03-05 11:33:01 +00:00
get: operations["get_log_level"];
2023-07-20 15:45:54 +00:00
/**
2023-08-22 05:15:00 +00:00
* Set Log Level
2023-07-20 15:45:54 +00:00
* @description Sets the log verbosity level
*/
2024-03-05 11:33:01 +00:00
post: operations["set_log_level"];
2023-07-20 15:45:54 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/invocation_cache": {
/**
* Clear Invocation Cache
* @description Clears the invocation cache
*/
2024-03-05 11:33:01 +00:00
delete: operations["clear_invocation_cache"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/invocation_cache/enable": {
/**
* Enable Invocation Cache
* @description Clears the invocation cache
*/
2024-03-05 11:33:01 +00:00
put: operations["enable_invocation_cache"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/invocation_cache/disable": {
/**
* Disable Invocation Cache
* @description Clears the invocation cache
*/
2024-03-05 11:33:01 +00:00
put: operations["disable_invocation_cache"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/app/invocation_cache/status": {
/**
* Get Invocation Cache Status
* @description Clears the invocation cache
*/
2024-03-05 11:33:01 +00:00
get: operations["get_invocation_cache_status"];
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/enqueue_batch": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Enqueue Batch
* @description Processes a batch and enqueues the output graphs for execution.
*/
2024-03-05 11:33:01 +00:00
post: operations["enqueue_batch"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/list": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* List Queue Items
* @description Gets all queue items (without graphs)
*/
2024-03-05 11:33:01 +00:00
get: operations["list_queue_items"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/processor/resume": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Resume
* @description Resumes session processor
*/
2024-03-05 11:33:01 +00:00
put: operations["resume"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/processor/pause": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Pause
* @description Pauses session processor
*/
2024-03-05 11:33:01 +00:00
put: operations["pause"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/cancel_by_batch_ids": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Cancel By Batch Ids
* @description Immediately cancels all queue items from the given batch ids
*/
2024-03-05 11:33:01 +00:00
put: operations["cancel_by_batch_ids"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/clear": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Clear
* @description Clears the queue entirely, immediately canceling the currently-executing session
*/
2024-03-05 11:33:01 +00:00
put: operations["clear"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/prune": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Prune
* @description Prunes all completed or errored queue items
*/
2024-03-05 11:33:01 +00:00
put: operations["prune"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/current": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Get Current Queue Item
* @description Gets the currently execution queue item
*/
2024-03-05 11:33:01 +00:00
get: operations["get_current_queue_item"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/next": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Get Next Queue Item
* @description Gets the next queue item, without executing it
*/
2024-03-05 11:33:01 +00:00
get: operations["get_next_queue_item"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/status": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Get Queue Status
* @description Gets the status of the session queue
*/
2024-03-05 11:33:01 +00:00
get: operations["get_queue_status"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/b/{batch_id}/status": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Get Batch Status
* @description Gets the status of the session queue
*/
2024-03-05 11:33:01 +00:00
get: operations["get_batch_status"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/i/{item_id}": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Get Queue Item
* @description Gets a queue item
*/
2024-03-05 11:33:01 +00:00
get: operations["get_queue_item"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/queue/{queue_id}/i/{item_id}/cancel": {
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Cancel Queue Item
* @description Deletes a queue item
*/
2024-03-05 11:33:01 +00:00
put: operations["cancel_queue_item"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/workflows/i/{workflow_id}": {
/**
* Get Workflow
* @description Gets a workflow
*/
2024-03-05 11:33:01 +00:00
get: operations["get_workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Delete Workflow
* @description Deletes a workflow
*/
2024-03-05 11:33:01 +00:00
delete: operations["delete_workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Update Workflow
* @description Updates a workflow
*/
2024-03-05 11:33:01 +00:00
patch: operations["update_workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
2024-03-05 11:33:01 +00:00
"/api/v1/workflows/": {
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* List Workflows
* @description Gets a page of workflows
*/
2024-03-05 11:33:01 +00:00
get: operations["list_workflows"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Create Workflow
* @description Creates a workflow
*/
2024-03-05 11:33:01 +00:00
post: operations["create_workflow"];
};
};
2024-03-05 11:33:01 +00:00
export type webhooks = Record<string, never>;
export type components = {
schemas: {
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** AddImagesToBoardResult */
AddImagesToBoardResult: {
/**
2023-08-22 05:15:00 +00:00
* Board Id
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description The id of the board the images were added to
*/
board_id: string;
/**
2023-08-22 05:15:00 +00:00
* Added Image Names
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description The image names that were added to the board
*/
2023-08-22 05:15:00 +00:00
added_image_names: string[];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Add Integers
* @description Adds two numbers
*/
AddInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* A
* @description The first number
* @default 0
*/
a?: number;
/**
2023-08-22 05:15:00 +00:00
* B
* @description The second number
* @default 0
*/
b?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default add
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "add";
};
2023-07-08 09:31:17 +00:00
/**
2023-08-22 05:15:00 +00:00
* AppConfig
2023-07-12 15:56:40 +00:00
* @description App Config Response
*/
AppConfig: {
2023-07-13 05:22:18 +00:00
/**
2023-08-22 05:15:00 +00:00
* Infill Methods
2023-07-13 05:22:18 +00:00
* @description List of available infill methods
*/
2023-08-22 05:15:00 +00:00
infill_methods: string[];
/**
2023-08-22 05:15:00 +00:00
* Upscaling Methods
* @description List of upscaling methods
*/
2024-03-05 11:33:01 +00:00
upscaling_methods: components["schemas"]["Upscaler"][];
/**
2023-08-22 05:15:00 +00:00
* Nsfw Methods
* @description List of NSFW checking methods
*/
2023-08-22 05:15:00 +00:00
nsfw_methods: string[];
/**
2023-08-22 05:15:00 +00:00
* Watermarking Methods
* @description List of invisible watermark methods
*/
2023-08-22 05:15:00 +00:00
watermarking_methods: string[];
2023-07-12 15:56:40 +00:00
};
2023-12-06 12:33:54 +00:00
/**
* AppDependencyVersions
* @description App depencency Versions Response
*/
AppDependencyVersions: {
/**
* Accelerate
* @description accelerate version
*/
accelerate: string;
/**
* Compel
* @description compel version
*/
compel: string;
/**
* Cuda
* @description CUDA version
*/
cuda: string | null;
/**
* Diffusers
* @description diffusers version
*/
diffusers: string;
/**
* Numpy
* @description Numpy version
*/
numpy: string;
/**
* Opencv
* @description OpenCV version
*/
opencv: string;
/**
* Onnx
* @description ONNX version
*/
onnx: string;
/**
* Pillow
* @description Pillow (PIL) version
*/
pillow: string;
/**
* Python
* @description Python version
*/
python: string;
/**
* Torch
* @description PyTorch version
*/
torch: string;
/**
* Torchvision
* @description PyTorch Vision version
*/
torchvision: string;
/**
* Transformers
* @description transformers version
*/
transformers: string;
/**
* Xformers
* @description xformers version
*/
xformers: string | null;
};
2023-07-12 15:56:40 +00:00
/**
2023-08-22 05:15:00 +00:00
* AppVersion
2023-07-08 09:31:17 +00:00
* @description App Version Response
*/
AppVersion: {
2023-07-13 05:22:18 +00:00
/**
2023-08-22 05:15:00 +00:00
* Version
2023-07-13 05:22:18 +00:00
* @description App version
*/
2023-07-08 09:31:17 +00:00
version: string;
};
/**
2024-01-22 22:28:24 +00:00
* BaseMetadata
* @description Adds typing data for discriminated union.
*/
2024-01-22 22:28:24 +00:00
BaseMetadata: {
/**
2024-01-22 22:28:24 +00:00
* Name
* @description model's name
*/
2024-01-22 22:28:24 +00:00
name: string;
/**
2024-01-22 22:28:24 +00:00
* Type
* @default basemetadata
* @constant
*/
2024-03-05 11:33:01 +00:00
type?: "basemetadata";
};
2024-02-15 11:15:21 +00:00
/**
* BaseModelType
* @description Base model type.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
BaseModelType: "any" | "sd-1" | "sd-2" | "sdxl" | "sdxl-refiner";
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** Batch */
Batch: {
/**
* Batch Id
* @description The ID of the batch
*/
batch_id?: string;
/**
* Data
* @description The batch data collection.
*/
2024-03-05 11:33:01 +00:00
data?: components["schemas"]["BatchDatum"][][] | null;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The graph to initialize the session with */
2024-03-05 11:33:01 +00:00
graph: components["schemas"]["Graph"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** @description The workflow to initialize the session with */
2024-03-05 11:33:01 +00:00
workflow?: components["schemas"]["WorkflowWithoutID"] | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Runs
* @description Int stating how many times to iterate through all possible batch indices
* @default 1
*/
runs: number;
};
/** BatchDatum */
BatchDatum: {
/**
* Node Path
* @description The node into which this batch data collection will be substituted.
*/
node_path: string;
/**
* Field Name
* @description The field into which this batch data collection will be substituted.
*/
field_name: string;
/**
* Items
* @description The list of items to substitute into the node/field.
*/
items?: (string | number)[];
};
/** BatchStatus */
BatchStatus: {
/**
* Queue Id
* @description The ID of the queue
*/
queue_id: string;
/**
* Batch Id
* @description The ID of the batch
*/
batch_id: string;
/**
* Pending
* @description Number of queue items with status 'pending'
*/
pending: number;
/**
* In Progress
* @description Number of queue items with status 'in_progress'
*/
in_progress: number;
/**
* Completed
* @description Number of queue items with status 'complete'
*/
completed: number;
/**
* Failed
* @description Number of queue items with status 'error'
*/
failed: number;
/**
* Canceled
* @description Number of queue items with status 'canceled'
*/
canceled: number;
/**
* Total
* @description Total number of queue items
*/
total: number;
};
2023-08-23 19:25:24 +00:00
/**
* Blank Image
* @description Creates a blank image and forwards it to the pipeline
*/
BlankImageInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2023-08-23 19:25:24 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-08-23 19:25:24 +00:00
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-08-23 19:25:24 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
2023-08-23 19:25:24 +00:00
/**
* Width
* @description The width of the image
* @default 512
*/
width?: number;
/**
* Height
* @description The height of the image
* @default 512
*/
height?: number;
/**
* Mode
* @description The mode of the image
* @default RGB
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
mode?: "RGB" | "RGBA";
2023-08-23 19:25:24 +00:00
/**
* @description The color of the image
* @default {
2023-11-16 00:21:28 +00:00
* "r": 0,
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* "g": 0,
2023-11-16 00:21:28 +00:00
* "b": 0,
* "a": 255
2023-08-23 19:25:24 +00:00
* }
*/
2024-03-05 11:33:01 +00:00
color?: components["schemas"]["ColorField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default blank_image
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "blank_image";
2023-08-23 19:25:24 +00:00
};
2023-08-22 05:15:00 +00:00
/**
* Blend Latents
* @description Blend two latents using a given alpha. Latents must have same size.
*/
BlendLatentsInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-08-22 05:15:00 +00:00
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-08-22 05:15:00 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents_a?: components["schemas"]["LatentsField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents_b?: components["schemas"]["LatentsField"];
2023-08-22 05:15:00 +00:00
/**
* Alpha
* @description Blending factor. 0.0 = use input A only, 1.0 = use input B only, 0.5 = 50% mix of input A and input B.
* @default 0.5
*/
alpha?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default lblend
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lblend";
2023-08-22 05:15:00 +00:00
};
/** BoardChanges */
BoardChanges: {
/**
2023-08-22 05:15:00 +00:00
* Board Name
* @description The board's new name.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
board_name?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Cover Image Name
* @description The name of the board's new cover image.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
cover_image_name?: string | null;
};
/**
2023-08-22 05:15:00 +00:00
* BoardDTO
* @description Deserialized board record with cover image URL and image count.
*/
BoardDTO: {
/**
2023-08-22 05:15:00 +00:00
* Board Id
* @description The unique ID of the board.
*/
board_id: string;
/**
2023-08-22 05:15:00 +00:00
* Board Name
* @description The name of the board.
*/
board_name: string;
/**
2023-08-22 05:15:00 +00:00
* Created At
* @description The created timestamp of the board.
*/
created_at: string;
/**
2023-08-22 05:15:00 +00:00
* Updated At
* @description The updated timestamp of the board.
*/
updated_at: string;
/**
2023-08-22 05:15:00 +00:00
* Deleted At
* @description The deleted timestamp of the board.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
deleted_at?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Cover Image Name
* @description The name of the board's cover image.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
cover_image_name: string | null;
/**
2023-08-22 05:15:00 +00:00
* Image Count
* @description The number of images in the board.
*/
image_count: number;
};
/**
* BoardField
* @description A board primitive field
*/
BoardField: {
/**
* Board Id
* @description The id of the board
*/
board_id: string;
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** Body_add_image_to_board */
Body_add_image_to_board: {
/**
2023-08-22 05:15:00 +00:00
* Board Id
* @description The id of the board to add to
*/
board_id: string;
/**
2023-08-22 05:15:00 +00:00
* Image Name
* @description The name of the image to add
*/
image_name: string;
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** Body_add_images_to_board */
Body_add_images_to_board: {
/**
2023-08-22 05:15:00 +00:00
* Board Id
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description The id of the board to add to
*/
board_id: string;
/**
2023-08-22 05:15:00 +00:00
* Image Names
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description The names of the images to add
*/
2023-08-22 05:15:00 +00:00
image_names: string[];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** Body_cancel_by_batch_ids */
Body_cancel_by_batch_ids: {
/**
* Batch Ids
* @description The list of batch_ids to cancel all queue items for
*/
batch_ids: string[];
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** Body_create_workflow */
Body_create_workflow: {
/** @description The workflow to create */
2024-03-05 11:33:01 +00:00
workflow: components["schemas"]["WorkflowWithoutID"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** Body_delete_images_from_list */
Body_delete_images_from_list: {
/**
2023-08-22 05:15:00 +00:00
* Image Names
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description The list of names of images to delete
*/
2023-08-22 05:15:00 +00:00
image_names: string[];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
/** Body_download */
Body_download: {
/**
* Source
* Format: uri
* @description download source
*/
source: string;
/**
* Dest
* @description download destination
*/
dest: string;
/**
* Priority
* @description queue priority
* @default 10
*/
priority?: number;
/**
* Access Token
* @description token for authorization to download
*/
access_token?: string | null;
};
/** Body_download_images_from_list */
Body_download_images_from_list: {
/**
* Image Names
* @description The list of names of images to download
*/
2024-02-20 11:15:14 +00:00
image_names?: string[] | null;
/**
* Board Id
2024-02-20 11:15:14 +00:00
* @description The board from which image should be downloaded
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
board_id?: string | null;
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** Body_enqueue_batch */
Body_enqueue_batch: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Batch to process */
2024-03-05 11:33:01 +00:00
batch: components["schemas"]["Batch"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Prepend
* @description Whether or not to prepend this batch in the queue
* @default false
*/
prepend?: boolean;
};
/** Body_parse_dynamicprompts */
Body_parse_dynamicprompts: {
/**
* Prompt
* @description The prompt to parse with dynamicprompts
*/
prompt: string;
/**
* Max Prompts
* @description The max number of prompts to generate
* @default 1000
*/
max_prompts?: number;
/**
* Combinatorial
* @description Whether to use the combinatorial generator
* @default true
*/
combinatorial?: boolean;
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** Body_remove_image_from_board */
Body_remove_image_from_board: {
/**
2023-08-22 05:15:00 +00:00
* Image Name
* @description The name of the image to remove
*/
image_name: string;
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** Body_remove_images_from_board */
Body_remove_images_from_board: {
/**
2023-08-22 05:15:00 +00:00
* Image Names
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description The names of the images to remove
*/
2023-08-22 05:15:00 +00:00
image_names: string[];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
/** Body_star_images_in_list */
Body_star_images_in_list: {
/**
2023-08-22 05:15:00 +00:00
* Image Names
* @description The list of names of images to star
*/
2023-08-22 05:15:00 +00:00
image_names: string[];
};
/** Body_unstar_images_in_list */
Body_unstar_images_in_list: {
/**
2023-08-22 05:15:00 +00:00
* Image Names
* @description The list of names of images to unstar
*/
2023-08-22 05:15:00 +00:00
image_names: string[];
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** Body_update_workflow */
Body_update_workflow: {
/** @description The updated workflow */
2024-03-05 11:33:01 +00:00
workflow: components["schemas"]["Workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/** Body_upload_image */
Body_upload_image: {
/**
2023-08-22 05:15:00 +00:00
* File
* Format: binary
*/
file: Blob;
};
/**
* Boolean Collection Primitive
* @description A collection of boolean primitive values
*/
BooleanCollectionInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Collection
* @description The collection of boolean values
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default []
*/
collection?: boolean[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default boolean_collection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "boolean_collection";
};
/**
2023-08-22 05:15:00 +00:00
* BooleanCollectionOutput
* @description Base class for nodes that output a collection of booleans
*/
BooleanCollectionOutput: {
/**
* Collection
* @description The output boolean collection
*/
collection: boolean[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default boolean_collection_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "boolean_collection_output";
};
/**
2023-08-22 05:15:00 +00:00
* Boolean Primitive
* @description A boolean primitive value
*/
BooleanInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Value
* @description The boolean value
* @default false
*/
value?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default boolean
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "boolean";
};
/**
2023-08-22 05:15:00 +00:00
* BooleanOutput
* @description Base class for nodes that output a single boolean
*/
BooleanOutput: {
/**
* Value
* @description The output boolean
*/
value: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default boolean_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "boolean_output";
};
2023-11-08 06:28:37 +00:00
/**
* CLIPOutput
* @description Base class for invocations that output a CLIP field
*/
CLIPOutput: {
/**
* CLIP
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip: components["schemas"]["ClipField"];
2023-11-08 06:28:37 +00:00
/**
* type
* @default clip_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "clip_output";
2023-11-08 06:28:37 +00:00
};
2023-11-16 00:21:28 +00:00
/**
* CLIPVisionDiffusersConfig
* @description Model config for ClipVision.
*/
CLIPVisionDiffusersConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
2023-11-16 00:21:28 +00:00
/**
* Description
* @description Model description
2023-11-16 00:21:28 +00:00
*/
description?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
2023-11-16 00:21:28 +00:00
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
2023-11-16 00:21:28 +00:00
*/
source_api_response?: string | null;
2024-02-15 11:15:21 +00:00
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
2024-02-15 11:15:21 +00:00
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2023-11-16 00:21:28 +00:00
/**
* Type
* @default clip_vision
* @constant
2023-11-16 00:21:28 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "clip_vision";
2023-09-14 15:19:52 +00:00
/**
* Format
* @constant
2023-09-14 15:19:52 +00:00
*/
format: "diffusers";
2023-09-14 15:19:52 +00:00
};
2023-09-05 02:17:23 +00:00
/**
* CV2 Infill
* @description Infills transparent areas of an image using OpenCV Inpainting
*/
CV2InfillInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2023-09-05 02:17:23 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to infill */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2023-09-05 02:17:23 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-09-05 02:17:23 +00:00
* @default infill_cv2
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-09-05 02:17:23 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "infill_cv2";
2023-09-05 02:17:23 +00:00
};
/**
* Calculate Image Tiles Even Split
* @description Calculate the coordinates and overlaps of tiles that cover a target image shape.
*/
CalculateImageTilesEvenSplitInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Image Width
* @description The image width, in pixels, to calculate tiles for.
* @default 1024
*/
image_width?: number;
/**
* Image Height
* @description The image height, in pixels, to calculate tiles for.
* @default 1024
*/
image_height?: number;
/**
* Num Tiles X
* @description Number of tiles to divide image into on the x axis
* @default 2
*/
num_tiles_x?: number;
/**
* Num Tiles Y
* @description Number of tiles to divide image into on the y axis
* @default 2
*/
num_tiles_y?: number;
/**
* Overlap
* @description The overlap, in pixels, between adjacent tiles.
* @default 128
*/
overlap?: number;
/**
* type
* @default calculate_image_tiles_even_split
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "calculate_image_tiles_even_split";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Calculate Image Tiles
* @description Calculate the coordinates and overlaps of tiles that cover a target image shape.
*/
CalculateImageTilesInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Image Width
* @description The image width, in pixels, to calculate tiles for.
* @default 1024
*/
image_width?: number;
/**
* Image Height
* @description The image height, in pixels, to calculate tiles for.
* @default 1024
*/
image_height?: number;
/**
* Tile Width
* @description The tile width, in pixels.
* @default 576
*/
tile_width?: number;
/**
* Tile Height
* @description The tile height, in pixels.
* @default 576
*/
tile_height?: number;
/**
* Overlap
* @description The target overlap, in pixels, between adjacent tiles. Adjacent tiles will overlap by at least this amount
* @default 128
*/
overlap?: number;
/**
* type
* @default calculate_image_tiles
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "calculate_image_tiles";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/**
* Calculate Image Tiles Minimum Overlap
* @description Calculate the coordinates and overlaps of tiles that cover a target image shape.
*/
CalculateImageTilesMinimumOverlapInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Image Width
* @description The image width, in pixels, to calculate tiles for.
* @default 1024
*/
image_width?: number;
/**
* Image Height
* @description The image height, in pixels, to calculate tiles for.
* @default 1024
*/
image_height?: number;
/**
* Tile Width
* @description The tile width, in pixels.
* @default 576
*/
tile_width?: number;
/**
* Tile Height
* @description The tile height, in pixels.
* @default 576
*/
tile_height?: number;
/**
* Min Overlap
* @description Minimum overlap between adjacent tiles, in pixels.
* @default 128
*/
min_overlap?: number;
/**
* type
* @default calculate_image_tiles_min_overlap
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "calculate_image_tiles_min_overlap";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** CalculateImageTilesOutput */
CalculateImageTilesOutput: {
/**
* Tiles
* @description The tiles coordinates that cover a particular image shape.
*/
2024-03-05 11:33:01 +00:00
tiles: components["schemas"]["Tile"][];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* type
* @default calculate_image_tiles_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "calculate_image_tiles_output";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* CancelByBatchIDsResult
* @description Result of canceling by list of batch ids
*/
CancelByBatchIDsResult: {
/**
* Canceled
* @description Number of queue items canceled
*/
canceled: number;
};
/**
2023-08-22 05:15:00 +00:00
* Canny Processor
* @description Canny edge detection for ControlNet
*/
CannyImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Low Threshold
* @description The low threshold of the Canny pixel gradient (0-255)
* @default 100
*/
low_threshold?: number;
/**
2023-08-22 05:15:00 +00:00
* High Threshold
* @description The high threshold of the Canny pixel gradient (0-255)
* @default 200
*/
high_threshold?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default canny_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "canny_image_processor";
};
2024-02-24 08:19:40 +00:00
/**
* Canvas Paste Back
* @description Combines two images by using the mask provided. Intended for use on the Unified Canvas.
*/
CanvasPasteBackInvocation: {
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
2024-02-24 08:19:40 +00:00
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2024-02-24 08:19:40 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The source image */
2024-03-05 11:33:01 +00:00
source_image?: components["schemas"]["ImageField"];
2024-02-24 08:19:40 +00:00
/** @description The target image */
2024-03-05 11:33:01 +00:00
target_image?: components["schemas"]["ImageField"];
2024-02-24 08:19:40 +00:00
/** @description The mask to use when pasting */
2024-03-05 11:33:01 +00:00
mask?: components["schemas"]["ImageField"];
2024-02-24 08:19:40 +00:00
/**
* Mask Blur
* @description The amount to blur the mask by
* @default 0
*/
mask_blur?: number;
/**
* type
* @default canvas_paste_back
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "canvas_paste_back";
2024-02-24 08:19:40 +00:00
};
/**
2023-12-06 12:33:54 +00:00
* Center Pad or Crop Image
* @description Pad or crop an image's sides from the center by specified pixels. Positive values are outside of the image.
*/
CenterPadCropInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The image to crop */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* Left
* @description Number of pixels to pad/crop from the left (negative values crop inwards, positive values pad outwards)
* @default 0
*/
left?: number;
/**
* Right
* @description Number of pixels to pad/crop from the right (negative values crop inwards, positive values pad outwards)
* @default 0
*/
right?: number;
/**
* Top
* @description Number of pixels to pad/crop from the top (negative values crop inwards, positive values pad outwards)
* @default 0
*/
top?: number;
/**
* Bottom
* @description Number of pixels to pad/crop from the bottom (negative values crop inwards, positive values pad outwards)
* @default 0
*/
bottom?: number;
/**
* type
2023-12-06 12:33:54 +00:00
* @default img_pad_crop
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_pad_crop";
};
2024-01-22 22:28:24 +00:00
/**
* CivitaiMetadata
* @description Extended metadata fields provided by Civitai.
*/
CivitaiMetadata: {
/**
* Name
* @description model's name
*/
name: string;
/**
* Files
* @description model files and their sizes
*/
2024-03-05 11:33:01 +00:00
files?: components["schemas"]["RemoteModelFile"][];
2024-01-22 22:28:24 +00:00
/**
* Type
* @default civitai
* @constant
*/
2024-03-05 11:33:01 +00:00
type?: "civitai";
2024-01-22 22:28:24 +00:00
/**
* Trigger Phrases
* @description Trigger phrases extracted from the API response
2024-01-22 22:28:24 +00:00
*/
trigger_phrases: string[];
2024-01-22 22:28:24 +00:00
/**
* Api Response
* @description Response from the Civitai API as stringified JSON
2024-01-22 22:28:24 +00:00
*/
api_response?: string | null;
2024-01-22 22:28:24 +00:00
};
/**
* CivitaiModelSource
* @description A Civitai version id, with optional variant and access token.
*/
CivitaiModelSource: {
/** Version Id */
version_id: number;
2024-03-05 11:33:01 +00:00
variant?: components["schemas"]["ModelRepoVariant"] | null;
2024-01-22 22:28:24 +00:00
/** Access Token */
access_token?: string | null;
/**
* Type
* @default civitai
* @constant
*/
2024-03-05 11:33:01 +00:00
type?: "civitai";
2024-01-22 22:28:24 +00:00
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* ClearResult
* @description Result of clearing the session queue
*/
ClearResult: {
/**
* Deleted
* @description Number of queue items deleted
*/
deleted: number;
};
/** ClipField */
ClipField: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Info to load tokenizer submodel */
2024-03-06 08:37:52 +00:00
tokenizer: components["schemas"]["ModelField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Info to load text_encoder submodel */
2024-03-06 08:37:52 +00:00
text_encoder: components["schemas"]["ModelField"];
2023-07-06 17:57:39 +00:00
/**
2023-08-22 05:15:00 +00:00
* Skipped Layers
2023-07-06 17:57:39 +00:00
* @description Number of skipped layers in text_encoder
*/
skipped_layers: number;
/**
2023-08-22 05:15:00 +00:00
* Loras
* @description Loras to apply on model loading
*/
2024-03-06 08:37:52 +00:00
loras: components["schemas"]["LoRAField"][];
};
2023-07-06 17:57:39 +00:00
/**
2023-08-22 05:15:00 +00:00
* CLIP Skip
2023-07-06 17:57:39 +00:00
* @description Skip layers in clip text_encoder model.
*/
ClipSkipInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-07-06 17:57:39 +00:00
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-07-06 17:57:39 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* CLIP
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip?: components["schemas"]["ClipField"];
2023-07-06 17:57:39 +00:00
/**
2023-08-22 05:15:00 +00:00
* Skipped Layers
* @description Number of layers to skip in text encoder
2023-07-06 17:57:39 +00:00
* @default 0
*/
skipped_layers?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default clip_skip
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "clip_skip";
2023-07-06 17:57:39 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* ClipSkipInvocationOutput
2023-07-06 17:57:39 +00:00
* @description Clip skip node output
*/
ClipSkipInvocationOutput: {
/**
* CLIP
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
2024-02-17 09:02:51 +00:00
* @default null
*/
2024-03-05 11:33:01 +00:00
clip: components["schemas"]["ClipField"] | null;
2023-07-06 17:57:39 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default clip_skip_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-07-06 17:57:39 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "clip_skip_output";
2023-07-06 17:57:39 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* CollectInvocation
* @description Collects values into a collection
*/
CollectInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Collection Item
* @description The item to collect (all inputs must be of the same type)
*/
item?: unknown;
/**
2023-08-22 05:15:00 +00:00
* Collection
* @description The collection, will be provided on execution
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default []
*/
2023-08-22 05:15:00 +00:00
collection?: unknown[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default collect
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "collect";
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** CollectInvocationOutput */
CollectInvocationOutput: {
/**
* Collection
* @description The collection of input items
*/
collection: unknown[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default collect_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "collect_output";
};
/**
2023-08-22 05:15:00 +00:00
* ColorCollectionOutput
* @description Base class for nodes that output a collection of colors
*/
ColorCollectionOutput: {
/**
* Collection
* @description The output colors
*/
2024-03-05 11:33:01 +00:00
collection: components["schemas"]["ColorField"][];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default color_collection_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "color_collection_output";
};
/**
2023-08-22 05:15:00 +00:00
* Color Correct
2023-08-11 18:15:59 +00:00
* @description Shifts the colors of a target image to match the reference image, optionally
* using a mask to only color-correct certain regions of the target image.
*/
ColorCorrectInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to color-correct */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Reference image for color-correction */
2024-03-05 11:33:01 +00:00
reference?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Mask to use when applying color-correction */
2024-03-05 11:33:01 +00:00
mask?: components["schemas"]["ImageField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Mask Blur Radius
* @description Mask blur radius
* @default 8
*/
mask_blur_radius?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default color_correct
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "color_correct";
};
/**
2023-08-22 05:15:00 +00:00
* ColorField
* @description A color primitive field
*/
ColorField: {
/**
2023-08-22 05:15:00 +00:00
* R
* @description The red component
*/
r: number;
/**
2023-08-22 05:15:00 +00:00
* G
* @description The green component
*/
g: number;
/**
2023-08-22 05:15:00 +00:00
* B
* @description The blue component
*/
b: number;
/**
2023-08-22 05:15:00 +00:00
* A
* @description The alpha component
*/
a: number;
};
/**
2023-08-22 05:15:00 +00:00
* Color Primitive
* @description A color primitive value
*/
ColorInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* @description The color value
* @default {
2023-11-16 00:21:28 +00:00
* "r": 0,
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* "g": 0,
2023-11-16 00:21:28 +00:00
* "b": 0,
* "a": 255
* }
*/
2024-03-05 11:33:01 +00:00
color?: components["schemas"]["ColorField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default color
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "color";
};
/**
* Color Map Processor
* @description Generates a color map from the provided image
*/
ColorMapImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* Color Map Tile Size
* @description Tile size
* @default 64
*/
color_map_tile_size?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default color_map_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "color_map_image_processor";
};
/**
2023-08-22 05:15:00 +00:00
* ColorOutput
* @description Base class for nodes that output a single color
*/
ColorOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The output color */
2024-03-05 11:33:01 +00:00
color: components["schemas"]["ColorField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default color_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "color_output";
};
/**
* Prompt
* @description Parse prompt using compel package to conditioning.
*/
CompelInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Prompt
* @description Prompt to be parsed by Compel to create a conditioning tensor
* @default
*/
prompt?: string;
/**
2023-08-22 05:15:00 +00:00
* CLIP
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip?: components["schemas"]["ClipField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default compel
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "compel";
};
/**
* Conditioning Collection Primitive
* @description A collection of conditioning tensor primitive values
*/
ConditioningCollectionInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Collection
* @description The collection of conditioning tensors
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default []
*/
2024-03-05 11:33:01 +00:00
collection?: components["schemas"]["ConditioningField"][];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default conditioning_collection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "conditioning_collection";
};
/**
2023-08-22 05:15:00 +00:00
* ConditioningCollectionOutput
* @description Base class for nodes that output a collection of conditioning tensors
*/
ConditioningCollectionOutput: {
/**
* Collection
* @description The output conditioning tensors
*/
2024-03-05 11:33:01 +00:00
collection: components["schemas"]["ConditioningField"][];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default conditioning_collection_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "conditioning_collection_output";
};
/**
2023-08-22 05:15:00 +00:00
* ConditioningField
* @description A conditioning tensor primitive value
*/
ConditioningField: {
/**
2023-08-22 05:15:00 +00:00
* Conditioning Name
* @description The name of conditioning tensor
*/
conditioning_name: string;
};
/**
2023-08-22 05:15:00 +00:00
* Conditioning Primitive
* @description A conditioning tensor primitive value
*/
ConditioningInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Conditioning tensor */
2024-03-05 11:33:01 +00:00
conditioning?: components["schemas"]["ConditioningField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default conditioning
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "conditioning";
};
/**
2023-08-22 05:15:00 +00:00
* ConditioningOutput
* @description Base class for nodes that output a single conditioning tensor
*/
ConditioningOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Conditioning tensor */
2024-03-05 11:33:01 +00:00
conditioning: components["schemas"]["ConditioningField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default conditioning_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "conditioning_output";
};
/**
2023-08-22 05:15:00 +00:00
* Content Shuffle Processor
* @description Applies content shuffle processing to image
*/
ContentShuffleImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* H
* @description Content shuffle `h` parameter
* @default 512
*/
h?: number;
/**
2023-08-22 05:15:00 +00:00
* W
* @description Content shuffle `w` parameter
* @default 512
*/
w?: number;
/**
2023-08-22 05:15:00 +00:00
* F
* @description Content shuffle `f` parameter
* @default 256
*/
f?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default content_shuffle_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "content_shuffle_image_processor";
};
/** ControlField */
ControlField: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The control image */
2024-03-05 11:33:01 +00:00
image: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The ControlNet model to use */
2024-03-06 08:37:52 +00:00
control_model: components["schemas"]["ModelField"];
/**
2023-08-22 05:15:00 +00:00
* Control Weight
* @description The weight given to the ControlNet
* @default 1
*/
2023-08-22 05:15:00 +00:00
control_weight?: number | number[];
/**
2023-08-22 05:15:00 +00:00
* Begin Step Percent
* @description When the ControlNet is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
2023-08-22 05:15:00 +00:00
* End Step Percent
* @description When the ControlNet is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
2023-06-25 04:04:16 +00:00
/**
2023-08-22 05:15:00 +00:00
* Control Mode
* @description The control mode to use
* @default balanced
2023-06-25 04:04:16 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
control_mode?: "balanced" | "more_prompt" | "more_control" | "unbalanced";
/**
2023-08-22 05:15:00 +00:00
* Resize Mode
* @description The resize mode to use
* @default just_resize
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
resize_mode?: "just_resize" | "crop_resize" | "fill_resize" | "just_resize_simple";
};
2023-11-16 00:21:28 +00:00
/**
* ControlNetCheckpointConfig
* @description Model config for ControlNet models (diffusers version).
*/
ControlNetCheckpointConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
2023-11-16 00:21:28 +00:00
/**
* Description
* @description Model description
2023-11-16 00:21:28 +00:00
*/
description?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
2023-11-16 00:21:28 +00:00
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
2023-11-16 00:21:28 +00:00
*/
source_api_response?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
2023-11-16 00:21:28 +00:00
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2024-02-15 11:15:21 +00:00
/**
* Format
* @default checkpoint
* @constant
2024-02-15 11:15:21 +00:00
*/
2024-03-05 11:33:01 +00:00
format: "checkpoint";
2023-11-16 00:21:28 +00:00
/**
* Config Path
* @description path to the checkpoint model config file
2023-11-16 00:21:28 +00:00
*/
config_path: string;
2024-02-15 11:15:21 +00:00
/**
* Converted At
* @description When this model was last converted to diffusers
2024-02-15 11:15:21 +00:00
*/
converted_at?: number | null;
2023-11-16 00:21:28 +00:00
/**
* Type
* @default controlnet
* @constant
2023-11-16 00:21:28 +00:00
*/
type: "controlnet";
2023-11-16 00:21:28 +00:00
};
/**
* ControlNetDiffusersConfig
* @description Model config for ControlNet models (diffusers version).
*/
ControlNetDiffusersConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
2023-11-16 00:21:28 +00:00
/**
* Description
* @description Model description
2023-11-16 00:21:28 +00:00
*/
description?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
2024-02-15 11:15:21 +00:00
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
2024-02-15 11:15:21 +00:00
*/
source_api_response?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
2023-11-16 00:21:28 +00:00
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2024-02-15 11:15:21 +00:00
/**
* Format
* @default diffusers
* @constant
2024-02-15 11:15:21 +00:00
*/
format: "diffusers";
2024-02-15 11:15:21 +00:00
/** @default */
repo_variant?: components["schemas"]["ModelRepoVariant"] | null;
/**
* Type
* @default controlnet
* @constant
*/
type: "controlnet";
2023-11-16 00:21:28 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* ControlNet
* @description Collects ControlNet info to pass to other nodes
*/
ControlNetInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The control image */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description ControlNet model to load */
2024-03-06 08:37:52 +00:00
control_model: components["schemas"]["ModelField"];
2023-07-13 05:22:18 +00:00
/**
2023-08-22 05:15:00 +00:00
* Control Weight
* @description The weight given to the ControlNet
* @default 1
*/
2023-08-22 05:15:00 +00:00
control_weight?: number | number[];
/**
2023-08-22 05:15:00 +00:00
* Begin Step Percent
* @description When the ControlNet is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
2023-08-22 05:15:00 +00:00
* End Step Percent
* @description When the ControlNet is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
2023-06-25 04:04:16 +00:00
/**
2023-08-22 05:15:00 +00:00
* Control Mode
* @description The control mode used
* @default balanced
2023-06-25 04:04:16 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
control_mode?: "balanced" | "more_prompt" | "more_control" | "unbalanced";
/**
2023-08-22 05:15:00 +00:00
* Resize Mode
* @description The resize mode used
* @default just_resize
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
resize_mode?: "just_resize" | "crop_resize" | "fill_resize" | "just_resize_simple";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default controlnet
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "controlnet";
};
2024-03-06 08:37:52 +00:00
/** ControlNetMetadataField */
ControlNetMetadataField: {
/** @description The control image */
image: components["schemas"]["ImageField"];
/** @description The ControlNet model to use */
control_model: components["schemas"]["ModelMetadataField"];
2023-07-08 09:31:45 +00:00
/**
2024-03-06 08:37:52 +00:00
* Control Weight
* @description The weight given to the ControlNet
* @default 1
2023-07-08 09:31:45 +00:00
*/
2024-03-06 08:37:52 +00:00
control_weight?: number | number[];
/**
* Begin Step Percent
* @description When the ControlNet is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
* End Step Percent
* @description When the ControlNet is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
/**
* Control Mode
* @description The control mode to use
* @default balanced
* @enum {string}
*/
control_mode?: "balanced" | "more_prompt" | "more_control" | "unbalanced";
/**
* Resize Mode
* @description The resize mode to use
* @default just_resize
* @enum {string}
*/
resize_mode?: "just_resize" | "crop_resize" | "fill_resize" | "just_resize_simple";
2023-07-08 09:31:45 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* ControlOutput
* @description node output for ControlNet info
*/
ControlOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description ControlNet(s) to apply */
2024-03-05 11:33:01 +00:00
control: components["schemas"]["ControlField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default control_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "control_output";
};
2023-07-12 15:15:09 +00:00
/**
* Core Metadata
* @description Collects core generation metadata into a MetadataField
2023-07-12 15:15:09 +00:00
*/
CoreMetadataInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
2023-07-12 15:15:09 +00:00
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
2023-07-12 15:15:09 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
2023-07-12 15:15:09 +00:00
*/
use_cache?: boolean;
2023-08-29 00:20:55 +00:00
/**
* Generation Mode
* @description The generation mode that output this image
2023-08-29 00:20:55 +00:00
*/
2024-03-05 11:33:01 +00:00
generation_mode?: ("txt2img" | "img2img" | "inpaint" | "outpaint" | "sdxl_txt2img" | "sdxl_img2img" | "sdxl_inpaint" | "sdxl_outpaint") | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Positive Prompt
2023-07-12 15:15:09 +00:00
* @description The positive prompt parameter
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
positive_prompt?: string | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Negative Prompt
2023-07-12 15:15:09 +00:00
* @description The negative prompt parameter
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
negative_prompt?: string | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Width
2023-07-12 15:15:09 +00:00
* @description The width parameter
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
width?: number | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Height
2023-07-12 15:15:09 +00:00
* @description The height parameter
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
height?: number | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Seed
2023-07-12 15:15:09 +00:00
* @description The seed used for noise generation
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
seed?: number | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Rand Device
2023-07-12 15:15:09 +00:00
* @description The device used for random number generation
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
rand_device?: string | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Cfg Scale
2023-07-12 15:15:09 +00:00
* @description The classifier-free guidance scale parameter
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
cfg_scale?: number | null;
/**
* Cfg Rescale Multiplier
* @description Rescale multiplier for CFG guidance, used for models trained with zero-terminal SNR
*/
cfg_rescale_multiplier?: number | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Steps
2023-07-12 15:15:09 +00:00
* @description The number of steps used for inference
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
steps?: number | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Scheduler
2023-07-12 15:15:09 +00:00
* @description The scheduler used for inference
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
scheduler?: string | null;
/**
* Seamless X
* @description Whether seamless tiling was used on the X axis
*/
seamless_x?: boolean | null;
/**
* Seamless Y
* @description Whether seamless tiling was used on the Y axis
*/
seamless_y?: boolean | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Clip Skip
2023-07-12 15:15:09 +00:00
* @description The number of skipped CLIP layers
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
clip_skip?: number | null;
/** @description The main model used for inference */
2024-03-06 08:37:52 +00:00
model?: components["schemas"]["ModelMetadataField"] | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Controlnets
2023-07-12 15:15:09 +00:00
* @description The ControlNets used for inference
*/
2024-03-06 08:37:52 +00:00
controlnets?: components["schemas"]["ControlNetMetadataField"][] | null;
/**
* Ipadapters
* @description The IP Adapters used for inference
*/
2024-03-05 11:33:01 +00:00
ipAdapters?: components["schemas"]["IPAdapterMetadataField"][] | null;
/**
* T2Iadapters
* @description The IP Adapters used for inference
*/
2024-03-06 08:37:52 +00:00
t2iAdapters?: components["schemas"]["T2IAdapterMetadataField"][] | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Loras
2023-07-12 15:15:09 +00:00
* @description The LoRAs used for inference
*/
2024-03-05 11:33:01 +00:00
loras?: components["schemas"]["LoRAMetadataField"][] | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Strength
2023-07-12 15:15:09 +00:00
* @description The strength used for latents-to-latents
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
strength?: number | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Init Image
2023-07-12 15:15:09 +00:00
* @description The name of the initial image
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
init_image?: string | null;
/** @description The VAE used for decoding, if the main model's default was not used */
2024-03-06 08:37:52 +00:00
vae?: components["schemas"]["ModelMetadataField"] | null;
/**
* Hrf Enabled
* @description Whether or not high resolution fix was enabled.
*/
hrf_enabled?: boolean | null;
/**
* Hrf Method
* @description The high resolution fix upscale method.
*/
hrf_method?: string | null;
/**
* Hrf Strength
* @description The high resolution fix img2img strength used in the upscale pass.
*/
hrf_strength?: number | null;
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Positive Style Prompt
* @description The positive style prompt parameter
2023-07-12 15:15:09 +00:00
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
positive_style_prompt?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Negative Style Prompt
* @description The negative style prompt parameter
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
negative_style_prompt?: string | null;
/** @description The SDXL Refiner model used */
2024-03-06 08:37:52 +00:00
refiner_model?: components["schemas"]["ModelMetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Refiner Cfg Scale
* @description The classifier-free guidance scale parameter used for the refiner
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
refiner_cfg_scale?: number | null;
/**
2023-08-22 05:15:00 +00:00
* Refiner Steps
* @description The number of steps used for the refiner
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
refiner_steps?: number | null;
/**
2023-08-22 05:15:00 +00:00
* Refiner Scheduler
* @description The scheduler used for the refiner
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
refiner_scheduler?: string | null;
/**
2023-09-02 01:36:07 +00:00
* Refiner Positive Aesthetic Score
* @description The aesthetic score used for the refiner
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
refiner_positive_aesthetic_score?: number | null;
/**
2023-09-02 01:36:07 +00:00
* Refiner Negative Aesthetic Score
* @description The aesthetic score used for the refiner
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
refiner_negative_aesthetic_score?: number | null;
/**
2023-08-22 05:15:00 +00:00
* Refiner Start
* @description The start value used for refiner denoising
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
refiner_start?: number | null;
/**
* type
* @default core_metadata
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "core_metadata";
[key: string]: unknown;
2023-07-12 15:15:09 +00:00
};
2023-08-29 00:20:55 +00:00
/**
* Create Denoise Mask
* @description Creates mask for denoising model run.
*/
CreateDenoiseMaskInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-08-29 00:20:55 +00:00
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-08-29 00:20:55 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description VAE */
2024-03-05 11:33:01 +00:00
vae?: components["schemas"]["VaeField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Image which will be masked */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"] | null;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The mask to use when pasting */
2024-03-05 11:33:01 +00:00
mask?: components["schemas"]["ImageField"];
2023-08-29 00:20:55 +00:00
/**
* Tiled
* @description Processing using overlapping tiles (reduce memory consumption)
* @default false
*/
tiled?: boolean;
/**
* Fp32
* @description Whether or not to use full float32 precision
2024-01-24 13:15:54 +00:00
* @default false
2023-08-29 00:20:55 +00:00
*/
fp32?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default create_denoise_mask
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "create_denoise_mask";
2023-08-29 00:20:55 +00:00
};
/**
* Create Gradient Mask
* @description Creates mask for denoising model run.
*/
CreateGradientMaskInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description Image which will be masked */
2024-03-05 11:33:01 +00:00
mask?: components["schemas"]["ImageField"];
/**
* Edge Radius
* @description How far to blur/expand the edges of the mask
* @default 16
*/
edge_radius?: number;
/**
* Coherence Mode
* @default Gaussian Blur
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
coherence_mode?: "Gaussian Blur" | "Box Blur" | "Staged";
/**
* Minimum Denoise
* @description Minimum denoise level for the coherence region
* @default 0
*/
minimum_denoise?: number;
/**
* type
* @default create_gradient_mask
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "create_gradient_mask";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Crop Latents
* @description Crops a latent-space tensor to a box specified in image-space. The box dimensions and coordinates must be
* divisible by the latent scale factor of 8.
*/
CropLatentsCoreInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents?: components["schemas"]["LatentsField"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* X
* @description The left x coordinate (in px) of the crop rectangle in image space. This value will be converted to a dimension in latent space.
*/
x?: number;
/**
* Y
* @description The top y coordinate (in px) of the crop rectangle in image space. This value will be converted to a dimension in latent space.
*/
y?: number;
/**
* Width
* @description The width (in px) of the crop rectangle in image space. This value will be converted to a dimension in latent space.
*/
width?: number;
/**
* Height
* @description The height (in px) of the crop rectangle in image space. This value will be converted to a dimension in latent space.
*/
height?: number;
/**
* type
* @default crop_latents
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "crop_latents";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** CursorPaginatedResults[SessionQueueItemDTO] */
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
CursorPaginatedResults_SessionQueueItemDTO_: {
/**
* Limit
* @description Limit of items to get
*/
limit: number;
/**
* Has More
* @description Whether there are more items available
*/
has_more: boolean;
/**
* Items
* @description Items
*/
2024-03-05 11:33:01 +00:00
items: components["schemas"]["SessionQueueItemDTO"][];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* OpenCV Inpaint
* @description Simple inpaint using opencv.
*/
CvInpaintInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to inpaint */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The mask to use when inpainting */
2024-03-05 11:33:01 +00:00
mask?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default cv_inpaint
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "cv_inpaint";
};
2024-02-09 20:14:15 +00:00
/**
2024-02-11 08:00:51 +00:00
* DW Openpose Image Processor
2024-02-09 20:14:15 +00:00
* @description Generates an openpose pose from an image using DWPose
*/
2024-02-11 08:00:51 +00:00
DWOpenposeImageProcessorInvocation: {
2024-02-13 04:02:30 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
2024-02-09 20:14:15 +00:00
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2024-02-09 20:14:15 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2024-02-09 20:14:15 +00:00
/**
* Draw Body
* @default true
*/
draw_body?: boolean;
/**
* Draw Face
* @default false
*/
draw_face?: boolean;
/**
* Draw Hands
* @default false
*/
draw_hands?: boolean;
/**
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
2024-02-09 20:14:15 +00:00
/**
* type
2024-02-11 08:00:51 +00:00
* @default dw_openpose_image_processor
2024-02-09 20:14:15 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "dw_openpose_image_processor";
2024-02-09 20:14:15 +00:00
};
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
/** DeleteBoardResult */
DeleteBoardResult: {
/**
2023-08-22 05:15:00 +00:00
* Board Id
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
* @description The id of the board that was deleted.
*/
board_id: string;
/**
2023-08-22 05:15:00 +00:00
* Deleted Board Images
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
* @description The image names of the board-images relationships that were deleted.
*/
2023-08-22 05:15:00 +00:00
deleted_board_images: string[];
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
/**
2023-08-22 05:15:00 +00:00
* Deleted Images
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
* @description The names of the images that were deleted.
*/
2023-08-22 05:15:00 +00:00
deleted_images: string[];
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** DeleteImagesFromListResult */
DeleteImagesFromListResult: {
/** Deleted Images */
2023-08-22 05:15:00 +00:00
deleted_images: string[];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Denoise Latents
* @description Denoises noisy latents to decodable images
*/
DenoiseLatentsInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Positive conditioning tensor */
2024-03-05 11:33:01 +00:00
positive_conditioning?: components["schemas"]["ConditioningField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Negative conditioning tensor */
2024-03-05 11:33:01 +00:00
negative_conditioning?: components["schemas"]["ConditioningField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Noise tensor */
2024-03-05 11:33:01 +00:00
noise?: components["schemas"]["LatentsField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Steps
* @description Number of steps to run
* @default 10
*/
steps?: number;
/**
2023-08-22 05:15:00 +00:00
* CFG Scale
* @description Classifier-Free Guidance scale
* @default 7.5
*/
2023-08-22 05:15:00 +00:00
cfg_scale?: number | number[];
/**
2023-08-22 05:15:00 +00:00
* Denoising Start
* @description When to start denoising, expressed a percentage of total steps
* @default 0
*/
denoising_start?: number;
/**
2023-08-22 05:15:00 +00:00
* Denoising End
* @description When to stop denoising, expressed a percentage of total steps
* @default 1
*/
denoising_end?: number;
/**
2023-08-22 05:15:00 +00:00
* Scheduler
* @description Scheduler to use during inference
* @default euler
2023-08-16 01:59:19 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
scheduler?: "ddim" | "ddpm" | "deis" | "lms" | "lms_k" | "pndm" | "heun" | "heun_k" | "euler" | "euler_k" | "euler_a" | "kdpm_2" | "kdpm_2_a" | "dpmpp_2s" | "dpmpp_2s_k" | "dpmpp_2m" | "dpmpp_2m_k" | "dpmpp_2m_sde" | "dpmpp_2m_sde_k" | "dpmpp_sde" | "dpmpp_sde_k" | "unipc" | "lcm";
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet?: components["schemas"]["UNetField"];
/** Control */
2024-03-05 11:33:01 +00:00
control?: components["schemas"]["ControlField"] | components["schemas"]["ControlField"][] | null;
2023-09-06 17:53:49 +00:00
/**
2023-09-12 23:09:22 +00:00
* IP-Adapter
2023-09-06 17:53:49 +00:00
* @description IP-Adapter to apply
*/
2024-03-05 11:33:01 +00:00
ip_adapter?: components["schemas"]["IPAdapterField"] | components["schemas"]["IPAdapterField"][] | null;
/**
* T2I-Adapter
* @description T2I-Adapter(s) to apply
*/
2024-03-05 11:33:01 +00:00
t2i_adapter?: components["schemas"]["T2IAdapterField"] | components["schemas"]["T2IAdapterField"][] | null;
/**
2024-01-22 22:28:24 +00:00
* CFG Rescale Multiplier
* @description Rescale multiplier for CFG guidance, used for models trained with zero-terminal SNR
* @default 0
*/
cfg_rescale_multiplier?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents?: components["schemas"]["LatentsField"] | null;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The mask to use for the operation */
2024-03-05 11:33:01 +00:00
denoise_mask?: components["schemas"]["DenoiseMaskField"] | null;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default denoise_latents
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "denoise_latents";
};
2023-08-29 00:20:55 +00:00
/**
* DenoiseMaskField
* @description An inpaint mask field
*/
DenoiseMaskField: {
/**
* Mask Name
* @description The name of the mask image
*/
mask_name: string;
/**
* Masked Latents Name
* @description The name of the masked image latents
2024-02-17 09:02:51 +00:00
* @default null
2023-08-29 00:20:55 +00:00
*/
2023-11-08 06:28:37 +00:00
masked_latents_name?: string | null;
/**
* Gradient
* @description Used for gradient inpainting
* @default false
*/
gradient?: boolean;
2023-08-29 00:20:55 +00:00
};
/**
* DenoiseMaskOutput
* @description Base class for nodes that output a single image
*/
DenoiseMaskOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Mask for denoise model run */
2024-03-05 11:33:01 +00:00
denoise_mask: components["schemas"]["DenoiseMaskField"];
2023-08-29 00:20:55 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-29 00:20:55 +00:00
* @default denoise_mask_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-08-29 00:20:55 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "denoise_mask_output";
2023-08-29 00:20:55 +00:00
};
2024-01-22 22:28:24 +00:00
/**
* Depth Anything Processor
* @description Generates a depth map based on the Depth Anything algorithm
*/
DepthAnythingImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
2024-01-22 22:28:24 +00:00
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2024-01-22 22:28:24 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2024-01-22 22:28:24 +00:00
/**
* Model Size
* @description The size of the depth model to use
2024-01-23 04:43:03 +00:00
* @default small
2024-01-22 22:28:24 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
model_size?: "large" | "base" | "small";
2024-01-23 04:43:03 +00:00
/**
* Resolution
* @description Pixel resolution for output image
* @default 512
*/
resolution?: number;
2024-01-22 22:28:24 +00:00
/**
* Offload
* @default false
*/
offload?: boolean;
/**
* type
* @default depth_anything_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "depth_anything_image_processor";
2024-01-22 22:28:24 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Divide Integers
* @description Divides two numbers
*/
DivideInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* A
* @description The first number
* @default 0
*/
a?: number;
/**
2023-08-22 05:15:00 +00:00
* B
* @description The second number
* @default 0
*/
b?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default div
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "div";
};
/**
* DownloadJob
* @description Class to monitor and control a model download request.
*/
DownloadJob: {
/**
* Source
* Format: uri
* @description Where to download from. Specific types specified in child classes.
*/
source: string;
/**
* Dest
* Format: path
* @description Destination of downloaded model on local disk; a directory or file path
*/
dest: string;
/**
* Access Token
* @description authorization token for protected resources
*/
access_token?: string | null;
/**
* Id
* @description Numeric ID of this job
* @default -1
*/
id?: number;
/**
* Priority
* @description Queue priority; lower values are higher priority
* @default 10
*/
priority?: number;
/**
* @description Status of the download
* @default waiting
*/
2024-03-05 11:33:01 +00:00
status?: components["schemas"]["DownloadJobStatus"];
/**
* Download Path
* @description Final location of downloaded file
*/
download_path?: string | null;
/**
* Job Started
* @description Timestamp for when the download job started
*/
job_started?: string | null;
/**
* Job Ended
* @description Timestamp for when the download job ende1d (completed or errored)
*/
job_ended?: string | null;
2024-01-22 22:28:24 +00:00
/**
* Content Type
* @description Content type of downloaded file
*/
content_type?: string | null;
/**
* Bytes
* @description Bytes downloaded so far
* @default 0
*/
bytes?: number;
/**
* Total Bytes
* @description Total file size (bytes)
* @default 0
*/
total_bytes?: number;
/**
* Error Type
* @description Name of exception that caused an error
*/
error_type?: string | null;
/**
* Error
* @description Traceback of the exception that caused an error
*/
error?: string | null;
};
/**
* DownloadJobStatus
* @description State of a download job.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
DownloadJobStatus: "waiting" | "running" | "completed" | "cancelled" | "error";
/**
2023-08-22 05:15:00 +00:00
* Dynamic Prompt
* @description Parses a prompt using adieyal/dynamicprompts' random or combinatorial generator
*/
DynamicPromptInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default false
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Prompt
* @description The prompt to parse with dynamicprompts
*/
prompt?: string;
/**
2023-08-22 05:15:00 +00:00
* Max Prompts
* @description The number of prompts to generate
* @default 1
*/
max_prompts?: number;
/**
2023-08-22 05:15:00 +00:00
* Combinatorial
* @description Whether to use the combinatorial generator
* @default false
*/
combinatorial?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default dynamic_prompt
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "dynamic_prompt";
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** DynamicPromptsResponse */
DynamicPromptsResponse: {
/** Prompts */
prompts: string[];
/** Error */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
error?: string | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
2023-07-17 11:08:53 +00:00
/**
2023-08-22 05:15:00 +00:00
* Upscale (RealESRGAN)
2023-07-17 11:08:53 +00:00
* @description Upscales an image using RealESRGAN.
*/
ESRGANInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2023-07-17 11:08:53 +00:00
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-07-17 11:08:53 +00:00
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-07-17 11:08:53 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The input image */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2023-07-17 11:08:53 +00:00
/**
2023-08-22 05:15:00 +00:00
* Model Name
* @description The Real-ESRGAN model to use
* @default RealESRGAN_x4plus.pth
2023-07-17 11:08:53 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
model_name?: "RealESRGAN_x4plus.pth" | "RealESRGAN_x4plus_anime_6B.pth" | "ESRGAN_SRx4_DF2KOST_official-ff704c30.pth" | "RealESRGAN_x2plus.pth";
2023-10-06 04:22:25 +00:00
/**
* Tile Size
* @description Tile size for tiled ESRGAN upscaling (0=tiling disabled)
* @default 400
*/
tile_size?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default esrgan
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "esrgan";
2023-07-17 11:08:53 +00:00
};
/** Edge */
Edge: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The connection for the edge's from node and field */
2024-03-05 11:33:01 +00:00
source: components["schemas"]["EdgeConnection"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The connection for the edge's to node and field */
2024-03-05 11:33:01 +00:00
destination: components["schemas"]["EdgeConnection"];
};
/** EdgeConnection */
EdgeConnection: {
/**
2023-08-22 05:15:00 +00:00
* Node Id
* @description The id of the node for this edge connection
*/
node_id: string;
/**
2023-08-22 05:15:00 +00:00
* Field
* @description The field for this connection
*/
field: string;
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** EnqueueBatchResult */
EnqueueBatchResult: {
/**
* Queue Id
* @description The ID of the queue
*/
queue_id: string;
/**
* Enqueued
* @description The total number of queue items enqueued
*/
enqueued: number;
/**
* Requested
* @description The total number of queue items requested to be enqueued
*/
requested: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The batch that was enqueued */
2024-03-05 11:33:01 +00:00
batch: components["schemas"]["Batch"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Priority
* @description The priority of the enqueued batch
*/
priority: number;
};
2024-02-13 04:02:30 +00:00
/** ExposedField */
ExposedField: {
/** Nodeid */
nodeId: string;
/** Fieldname */
fieldName: string;
};
2024-02-09 20:14:15 +00:00
/**
2024-02-13 04:02:30 +00:00
* FaceIdentifier
* @description Outputs an image with detected face IDs printed on each face. For use with other FaceTools.
2024-02-09 20:14:15 +00:00
*/
2024-02-13 04:02:30 +00:00
FaceIdentifierInvocation: {
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
2024-02-09 20:14:15 +00:00
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2024-02-09 20:14:15 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
2024-02-13 04:02:30 +00:00
/** @description Image to face detect */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2024-02-09 20:14:15 +00:00
/**
2024-02-13 04:02:30 +00:00
* Minimum Confidence
* @description Minimum confidence for face detection (lower if detection is failing)
* @default 0.5
*/
minimum_confidence?: number;
/**
* Chunk
* @description Whether to bypass full image face detection and default to image chunking. Chunking will occur if no faces are found in the full image.
* @default false
*/
chunk?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default face_identifier
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "face_identifier";
};
/**
* FaceMask
* @description Face mask creation using mediapipe face detection
*/
FaceMaskInvocation: {
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Image to face detect */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* Face Ids
* @description Comma-separated list of face ids to mask eg '0,2,7'. Numbered from 0. Leave empty to mask all. Find face IDs with FaceIdentifier node.
* @default
*/
face_ids?: string;
/**
* Minimum Confidence
* @description Minimum confidence for face detection (lower if detection is failing)
* @default 0.5
*/
minimum_confidence?: number;
/**
* X Offset
* @description Offset for the X-axis of the face mask
* @default 0
*/
x_offset?: number;
/**
* Y Offset
* @description Offset for the Y-axis of the face mask
* @default 0
*/
y_offset?: number;
/**
* Chunk
* @description Whether to bypass full image face detection and default to image chunking. Chunking will occur if no faces are found in the full image.
* @default false
*/
chunk?: boolean;
/**
* Invert Mask
* @description Toggle to invert the mask
* @default false
*/
invert_mask?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default face_mask_detection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "face_mask_detection";
};
/**
* FaceMaskOutput
* @description Base class for FaceMask output
*/
FaceMaskOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The output image */
2024-03-05 11:33:01 +00:00
image: components["schemas"]["ImageField"];
/**
* Width
* @description The width of the image in pixels
*/
width: number;
/**
* Height
* @description The height of the image in pixels
*/
height: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default face_mask_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "face_mask_output";
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The output mask */
2024-03-05 11:33:01 +00:00
mask: components["schemas"]["ImageField"];
};
/**
* FaceOff
* @description Bound, extract, and mask a face from an image using MediaPipe detection
*/
FaceOffInvocation: {
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Image for face detection */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* Face Id
* @description The face ID to process, numbered from 0. Multiple faces not supported. Find a face's ID with FaceIdentifier node.
* @default 0
*/
face_id?: number;
/**
* Minimum Confidence
* @description Minimum confidence for face detection (lower if detection is failing)
* @default 0.5
*/
minimum_confidence?: number;
/**
* X Offset
* @description X-axis offset of the mask
* @default 0
*/
x_offset?: number;
/**
* Y Offset
* @description Y-axis offset of the mask
* @default 0
*/
y_offset?: number;
/**
* Padding
* @description All-axis padding around the mask in pixels
* @default 0
*/
padding?: number;
/**
* Chunk
* @description Whether to bypass full image face detection and default to image chunking. Chunking will occur if no faces are found in the full image.
* @default false
*/
chunk?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default face_off
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "face_off";
};
/**
* FaceOffOutput
* @description Base class for FaceOff Output
*/
FaceOffOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The output image */
2024-03-05 11:33:01 +00:00
image: components["schemas"]["ImageField"];
/**
* Width
* @description The width of the image in pixels
*/
width: number;
/**
* Height
* @description The height of the image in pixels
*/
height: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default face_off_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "face_off_output";
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The output mask */
2024-03-05 11:33:01 +00:00
mask: components["schemas"]["ImageField"];
/**
* X
* @description The x coordinate of the bounding box's left side
*/
x: number;
/**
* Y
* @description The y coordinate of the bounding box's top side
*/
y: number;
};
/**
* Float Collection Primitive
* @description A collection of float primitive values
*/
FloatCollectionInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Collection
* @description The collection of float values
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default []
*/
collection?: number[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default float_collection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "float_collection";
};
/**
2023-08-22 05:15:00 +00:00
* FloatCollectionOutput
* @description Base class for nodes that output a collection of floats
*/
FloatCollectionOutput: {
/**
* Collection
* @description The float collection
*/
collection: number[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default float_collection_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "float_collection_output";
};
/**
2023-08-22 05:15:00 +00:00
* Float Primitive
* @description A float primitive value
*/
FloatInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Value
* @description The float value
* @default 0
*/
value?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default float
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "float";
};
/**
2023-08-22 05:15:00 +00:00
* Float Range
* @description Creates a range
*/
FloatLinearRangeInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Start
* @description The first value of the range
* @default 5
*/
start?: number;
/**
2023-08-22 05:15:00 +00:00
* Stop
* @description The last value of the range
* @default 10
*/
stop?: number;
/**
2023-08-22 05:15:00 +00:00
* Steps
* @description number of values to interpolate over (including start and stop)
* @default 30
*/
steps?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default float_range
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "float_range";
};
/**
* Float Math
* @description Performs floating point math.
*/
FloatMathInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Operation
* @description The operation to perform
* @default ADD
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
operation?: "ADD" | "SUB" | "MUL" | "DIV" | "EXP" | "ABS" | "SQRT" | "MIN" | "MAX";
/**
* A
* @description The first number
2023-11-08 06:28:37 +00:00
* @default 1
*/
a?: number;
/**
* B
* @description The second number
2023-11-08 06:28:37 +00:00
* @default 1
*/
b?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default float_math
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "float_math";
};
/**
2023-08-22 05:15:00 +00:00
* FloatOutput
* @description Base class for nodes that output a single float
*/
FloatOutput: {
/**
* Value
* @description The output float
*/
value: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default float_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "float_output";
};
/**
* Float To Integer
* @description Rounds a float number to (a multiple of) an integer.
*/
FloatToIntegerInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Value
* @description The value to round
* @default 0
*/
value?: number;
/**
* Multiple of
* @description The multiple to round to
* @default 1
*/
multiple?: number;
/**
* Method
* @description The method to use for rounding
* @default Nearest
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
method?: "Nearest" | "Floor" | "Ceiling" | "Truncate";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default float_to_int
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "float_to_int";
};
2024-02-24 07:47:06 +00:00
/** FoundModel */
FoundModel: {
/**
* Path
* @description Path to the model
*/
path: string;
/**
* Is Installed
* @description Whether or not the model is already installed
*/
is_installed: boolean;
};
2023-11-08 06:28:37 +00:00
/**
* FreeUConfig
* @description Configuration for the FreeU hyperparameters.
* - https://huggingface.co/docs/diffusers/main/en/using-diffusers/freeu
* - https://github.com/ChenyangSi/FreeU
*/
FreeUConfig: {
/**
* S1
* @description Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to mitigate the "oversmoothing effect" in the enhanced denoising process.
*/
s1: number;
/**
* S2
* @description Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to mitigate the "oversmoothing effect" in the enhanced denoising process.
*/
s2: number;
/**
* B1
* @description Scaling factor for stage 1 to amplify the contributions of backbone features.
*/
b1: number;
/**
* B2
* @description Scaling factor for stage 2 to amplify the contributions of backbone features.
*/
b2: number;
};
/**
* FreeU
* @description Applies FreeU to the UNet. Suggested values (b1/b2/s1/s2):
*
* SD1.5: 1.2/1.4/0.9/0.2,
* SD2: 1.1/1.2/0.9/0.2,
* SDXL: 1.1/1.2/0.6/0.4,
*/
FreeUInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet?: components["schemas"]["UNetField"];
2023-11-08 06:28:37 +00:00
/**
* B1
* @description Scaling factor for stage 1 to amplify the contributions of backbone features.
* @default 1.2
*/
b1?: number;
/**
* B2
* @description Scaling factor for stage 2 to amplify the contributions of backbone features.
* @default 1.4
*/
b2?: number;
/**
* S1
* @description Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to mitigate the "oversmoothing effect" in the enhanced denoising process.
* @default 0.9
*/
s1?: number;
/**
* S2
* @description Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to mitigate the "oversmoothing effect" in the enhanced denoising process.
* @default 0.2
*/
s2?: number;
/**
* type
* @default freeu
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "freeu";
};
/**
* GradientMaskOutput
* @description Outputs a denoise mask and an image representing the total gradient of the mask.
*/
GradientMaskOutput: {
/** @description Mask for denoise model run */
denoise_mask: components["schemas"]["DenoiseMaskField"];
/** @description Image representing the total gradient area of the mask. For paste-back purposes. */
expanded_mask_area: components["schemas"]["ImageField"];
/**
* type
* @default gradient_mask_output
* @constant
*/
type: "gradient_mask_output";
};
/** Graph */
Graph: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this graph
*/
2024-02-20 11:15:14 +00:00
id?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Nodes
* @description The nodes in this graph
*/
2024-02-17 09:02:51 +00:00
nodes: {
2024-03-06 08:37:52 +00:00
[key: string]: components["schemas"]["ContentShuffleImageProcessorInvocation"] | components["schemas"]["RoundInvocation"] | components["schemas"]["ImageInvocation"] | components["schemas"]["SDXLModelLoaderInvocation"] | components["schemas"]["CoreMetadataInvocation"] | components["schemas"]["MultiplyInvocation"] | components["schemas"]["ConditioningInvocation"] | components["schemas"]["ImagePasteInvocation"] | components["schemas"]["LoraLoaderInvocation"] | components["schemas"]["RandomFloatInvocation"] | components["schemas"]["CreateDenoiseMaskInvocation"] | components["schemas"]["DynamicPromptInvocation"] | components["schemas"]["MetadataItemInvocation"] | components["schemas"]["MediapipeFaceProcessorInvocation"] | components["schemas"]["IdealSizeInvocation"] | components["schemas"]["InfillTileInvocation"] | components["schemas"]["LatentsCollectionInvocation"] | components["schemas"]["ImageCropInvocation"] | components["schemas"]["DWOpenposeImageProcessorInvocation"] | components["schemas"]["DivideInvocation"] | components["schemas"]["NormalbaeImageProcessorInvocation"] | components["schemas"]["FaceIdentifierInvocation"] | components["schemas"]["CanvasPasteBackInvocation"] | components["schemas"]["ImageChannelInvocation"] | components["schemas"]["FaceOffInvocation"] | components["schemas"]["MergeTilesToImageInvocation"] | components["schemas"]["FaceMaskInvocation"] | components["schemas"]["ImageScaleInvocation"] | components["schemas"]["ScaleLatentsInvocation"] | components["schemas"]["CreateGradientMaskInvocation"] | components["schemas"]["CenterPadCropInvocation"] | components["schemas"]["SubtractInvocation"] | components["schemas"]["BlankImageInvocation"] | components["schemas"]["FloatLinearRangeInvocation"] | components["schemas"]["FreeUInvocation"] | components["schemas"]["ImageResizeInvocation"] | components["schemas"]["CV2InfillInvocation"] | components["schemas"]["TileToPropertiesInvocation"] | components["schemas"]["ImageMultiplyInvocation"] | components["schemas"]["ZoeDepthImageProcessorInvocation"] | components["schemas"]["SaveImageInvocation"] | components["schemas"]["FloatInvocation"] | components["schemas"]["ImageHueAdjustmentInvocation"] | components["schemas"]["PidiImageProcessorInvocation"] | components["schemas"]["ConditioningCollectionInvocation"] | components["schemas"]["LeresImageProcessorInvocation"] | components["schemas"]["ImageInverseLerpInvocation"] | components["schemas"]["RangeInvocation"] | components["schemas"]["UnsharpMaskInvocation"] | components["schemas"]["BlendLatentsInvocation"] | components["schemas"]["PromptsFromFileInvocation"] | components["schemas"]["SDXLCompelPromptInvocation"] | components["schemas"]["StringInvocation"] | components["schemas"]["CropLatentsCoreInvocation"] | components["schemas"]["RangeOfSizeInvocation"] | components["schemas"]["LineartAnimeImageProcessorInvocation"] | components["schemas"]["IntegerCollectionInvocation"] | components["schemas"]["NoiseInvocation"] | components["schemas"]["SeamlessModeInvocation"] | components["schemas"]["SDXLLoraLoaderInvocation"] | components["schemas"]["StringReplaceInvocation"] | components["schemas"]["ShowImageInvocation"] | components["schemas"]["InfillColorInvocation"] | components["schemas"]["SegmentAnythingProcessorInvocation"] | components["schemas"]["BooleanInvocation"] | components["schemas"]["CompelInvocation"] | components["schemas"]["StringJoinInvocation"] | components["schemas"]["CalculateImageTilesInvocation"] | components["schemas"]["RandomRangeInvocation"] | components["schemas"]["ColorInvocation"] | components["schemas"]["CalculateImageTilesMinimumOverlapInvocation"] | components["schemas"]["ImageCollectionInvocation"] | components["schemas"]["ImageChannelMultiplyInvocation"] | components["schemas"]["MaskFromAlphaInvocation"] | components["schemas"]["ImageLerpInvocation"] | components["schemas"]["AddInvocation"] | components["schemas"]["DepthAnythingImageProcessorInvocation"] | components["schemas"]["StringJoinThreeInvocation"] | components["schemas"]["SDXLRefinerModelLoaderInvocation"] | components["
2023-08-16 01:59:19 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Edges
* @description The connections between nodes and their fields in this graph
*/
2024-03-05 11:33:01 +00:00
edges: components["schemas"]["Edge"][];
};
/**
2023-08-22 05:15:00 +00:00
* GraphExecutionState
* @description Tracks the state of a graph execution
*/
GraphExecutionState: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of the execution state
*/
id: string;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The graph being executed */
2024-03-05 11:33:01 +00:00
graph: components["schemas"]["Graph"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The expanded graph of activated and executed nodes */
2024-03-05 11:33:01 +00:00
execution_graph: components["schemas"]["Graph"];
/**
2023-08-22 05:15:00 +00:00
* Executed
* @description The set of node ids that have been executed
*/
2023-08-22 05:15:00 +00:00
executed: string[];
/**
2023-08-22 05:15:00 +00:00
* Executed History
* @description The list of node ids that have been executed, in order of execution
*/
2023-08-22 05:15:00 +00:00
executed_history: string[];
/**
2023-08-22 05:15:00 +00:00
* Results
* @description The results of node executions
*/
results: {
2024-03-06 08:37:52 +00:00
[key: string]: components["schemas"]["StringCollectionOutput"] | components["schemas"]["LatentsCollectionOutput"] | components["schemas"]["FloatCollectionOutput"] | components["schemas"]["ClipSkipInvocationOutput"] | components["schemas"]["NoiseOutput"] | components["schemas"]["PairTileImageOutput"] | components["schemas"]["BooleanCollectionOutput"] | components["schemas"]["BooleanOutput"] | components["schemas"]["VAEOutput"] | components["schemas"]["CalculateImageTilesOutput"] | components["schemas"]["StringOutput"] | components["schemas"]["SeamlessModeOutput"] | components["schemas"]["UNetOutput"] | components["schemas"]["LatentsOutput"] | components["schemas"]["FloatOutput"] | components["schemas"]["IntegerOutput"] | components["schemas"]["CollectInvocationOutput"] | components["schemas"]["ImageOutput"] | components["schemas"]["ConditioningCollectionOutput"] | components["schemas"]["FaceMaskOutput"] | components["schemas"]["TileToPropertiesOutput"] | components["schemas"]["ColorOutput"] | components["schemas"]["T2IAdapterOutput"] | components["schemas"]["ColorCollectionOutput"] | components["schemas"]["GradientMaskOutput"] | components["schemas"]["SDXLModelLoaderOutput"] | components["schemas"]["String2Output"] | components["schemas"]["LoraLoaderOutput"] | components["schemas"]["StringPosNegOutput"] | components["schemas"]["MetadataOutput"] | components["schemas"]["ModelLoaderOutput"] | components["schemas"]["SDXLRefinerModelLoaderOutput"] | components["schemas"]["MetadataItemOutput"] | components["schemas"]["ControlOutput"] | components["schemas"]["SDXLLoraLoaderOutput"] | components["schemas"]["IPAdapterOutput"] | components["schemas"]["ImageCollectionOutput"] | components["schemas"]["CLIPOutput"] | components["schemas"]["FaceOffOutput"] | components["schemas"]["IntegerCollectionOutput"] | components["schemas"]["SchedulerOutput"] | components["schemas"]["ConditioningOutput"] | components["schemas"]["DenoiseMaskOutput"] | components["schemas"]["IterateInvocationOutput"] | components["schemas"]["IdealSizeOutput"];
2023-08-16 01:59:19 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Errors
* @description Errors raised when executing nodes
*/
errors: {
2023-08-22 05:15:00 +00:00
[key: string]: string;
};
/**
2023-08-22 05:15:00 +00:00
* Prepared Source Mapping
* @description The map of prepared nodes to original graph nodes
*/
prepared_source_mapping: {
2023-08-22 05:15:00 +00:00
[key: string]: string;
};
/**
2023-08-22 05:15:00 +00:00
* Source Prepared Mapping
* @description The map of original graph nodes to prepared nodes
*/
source_prepared_mapping: {
2023-08-22 05:15:00 +00:00
[key: string]: string[];
};
};
/**
* HFModelSource
2024-01-22 22:28:24 +00:00
* @description A HuggingFace repo_id with optional variant, sub-folder and access token.
* Note that the variant option, if not provided to the constructor, will default to fp16, which is
* what people (almost) always want.
*/
HFModelSource: {
/** Repo Id */
repo_id: string;
2024-01-22 22:28:24 +00:00
/** @default fp16 */
2024-03-05 11:33:01 +00:00
variant?: components["schemas"]["ModelRepoVariant"] | null;
/** Subfolder */
subfolder?: string | null;
/** Access Token */
access_token?: string | null;
/**
* Type
* @default hf
* @constant
*/
2024-03-05 11:33:01 +00:00
type?: "hf";
};
/** HTTPValidationError */
HTTPValidationError: {
/** Detail */
2024-03-05 11:33:01 +00:00
detail?: components["schemas"]["ValidationError"][];
};
/**
2023-08-22 05:15:00 +00:00
* HED (softedge) Processor
* @description Applies HED edge detection to image
*/
HedImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Scribble
* @description Whether or not to use scribble mode
* @default false
*/
scribble?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default hed_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "hed_image_processor";
};
2024-01-22 22:28:24 +00:00
/**
* HuggingFaceMetadata
* @description Extended metadata fields provided by HuggingFace.
*/
HuggingFaceMetadata: {
/**
* Name
* @description model's name
*/
name: string;
/**
* Files
* @description model files and their sizes
2024-01-22 22:28:24 +00:00
*/
2024-03-05 11:33:01 +00:00
files?: components["schemas"]["RemoteModelFile"][];
2024-01-22 22:28:24 +00:00
/**
* Type
* @default huggingface
* @constant
*/
2024-03-05 11:33:01 +00:00
type?: "huggingface";
2024-01-22 22:28:24 +00:00
/**
* Id
* @description The HF model id
2024-01-22 22:28:24 +00:00
*/
id: string;
/**
* Api Response
* @description Response from the HF API as stringified JSON
2024-01-22 22:28:24 +00:00
*/
api_response?: string | null;
2024-01-22 22:28:24 +00:00
};
2023-11-16 00:21:28 +00:00
/**
* IPAdapterConfig
* @description Model config for IP Adaptor format models.
*/
IPAdapterConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
2023-11-16 00:21:28 +00:00
/**
* Description
* @description Model description
2023-11-16 00:21:28 +00:00
*/
description?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
2024-02-15 11:15:21 +00:00
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
2024-02-15 11:15:21 +00:00
*/
source_api_response?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
2023-11-16 00:21:28 +00:00
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2024-02-15 11:15:21 +00:00
/**
* Type
* @default ip_adapter
* @constant
2024-02-15 11:15:21 +00:00
*/
type: "ip_adapter";
2024-02-15 11:15:21 +00:00
/** Image Encoder Model Id */
image_encoder_model_id: string;
/**
* Format
* @constant
*/
format: "invokeai";
2023-11-16 00:21:28 +00:00
};
2023-09-06 17:53:49 +00:00
/** IPAdapterField */
IPAdapterField: {
/**
* Image
* @description The IP-Adapter image prompt(s).
*/
2024-03-05 11:33:01 +00:00
image: components["schemas"]["ImageField"] | components["schemas"]["ImageField"][];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The IP-Adapter model to use. */
2024-03-06 08:37:52 +00:00
ip_adapter_model: components["schemas"]["ModelField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The name of the CLIP image encoder model. */
2024-03-06 08:37:52 +00:00
image_encoder_model: components["schemas"]["ModelField"];
2023-09-06 17:53:49 +00:00
/**
* Weight
* @description The weight given to the ControlNet
2023-09-06 17:53:49 +00:00
* @default 1
*/
weight?: number | number[];
/**
* Begin Step Percent
* @description When the IP-Adapter is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
* End Step Percent
* @description When the IP-Adapter is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
2023-09-06 17:53:49 +00:00
};
2023-09-05 03:23:16 +00:00
/**
* IP-Adapter
2023-09-06 17:53:49 +00:00
* @description Collects IP-Adapter info to pass to other nodes.
2023-09-05 03:23:16 +00:00
*/
IPAdapterInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Image
* @description The IP-Adapter image prompt(s).
*/
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"] | components["schemas"]["ImageField"][];
2023-09-05 03:23:16 +00:00
/**
2023-09-12 23:09:22 +00:00
* IP-Adapter Model
* @description The IP-Adapter model.
2023-09-05 03:23:16 +00:00
*/
2024-03-06 08:37:52 +00:00
ip_adapter_model: components["schemas"]["ModelField"];
2023-09-05 03:23:16 +00:00
/**
2023-09-06 17:53:49 +00:00
* Weight
* @description The weight given to the IP-Adapter
2023-09-05 03:23:16 +00:00
* @default 1
*/
weight?: number | number[];
/**
* Begin Step Percent
* @description When the IP-Adapter is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
* End Step Percent
* @description When the IP-Adapter is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
2023-09-05 03:23:16 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-09-06 17:53:49 +00:00
* @default ip_adapter
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-09-05 03:23:16 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "ip_adapter";
2023-09-06 17:53:49 +00:00
};
/**
* IPAdapterMetadataField
* @description IP Adapter Field, minus the CLIP Vision Encoder model
*/
IPAdapterMetadataField: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The IP-Adapter image prompt. */
2024-03-05 11:33:01 +00:00
image: components["schemas"]["ImageField"];
/** @description The IP-Adapter model. */
2024-03-06 08:37:52 +00:00
ip_adapter_model: components["schemas"]["ModelMetadataField"];
/**
* Weight
* @description The weight given to the IP-Adapter
*/
weight: number | number[];
/**
* Begin Step Percent
* @description When the IP-Adapter is first applied (% of total steps)
*/
begin_step_percent: number;
/**
* End Step Percent
* @description When the IP-Adapter is last applied (% of total steps)
*/
end_step_percent: number;
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** IPAdapterOutput */
2023-09-06 17:53:49 +00:00
IPAdapterOutput: {
/**
2023-09-12 23:09:22 +00:00
* IP-Adapter
2023-09-06 17:53:49 +00:00
* @description IP-Adapter to apply
*/
2024-03-05 11:33:01 +00:00
ip_adapter: components["schemas"]["IPAdapterField"];
2023-09-06 17:53:49 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-09-06 17:53:49 +00:00
* @default ip_adapter_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-09-06 17:53:49 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "ip_adapter_output";
};
/**
* Ideal Size
* @description Calculates the ideal size for generation to avoid duplication
*/
IdealSizeInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Width
2024-01-22 22:28:24 +00:00
* @description Final image width
* @default 1024
*/
width?: number;
/**
* Height
2024-01-22 22:28:24 +00:00
* @description Final image height
* @default 576
*/
height?: number;
2024-01-22 22:28:24 +00:00
/** @description UNet (scheduler, LoRAs) */
2024-03-05 11:33:01 +00:00
unet?: components["schemas"]["UNetField"];
/**
* Multiplier
2024-01-22 22:28:24 +00:00
* @description Amount to multiply the model's dimensions by when calculating the ideal size (may result in initial generation artifacts if too large)
* @default 1
*/
multiplier?: number;
/**
* type
* @default ideal_size
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "ideal_size";
};
/**
* IdealSizeOutput
* @description Base class for invocations that output an image
*/
IdealSizeOutput: {
/**
* Width
2024-01-22 22:28:24 +00:00
* @description The ideal width of the image (in pixels)
*/
width: number;
/**
* Height
2024-01-22 22:28:24 +00:00
* @description The ideal height of the image (in pixels)
*/
height: number;
/**
* type
* @default ideal_size_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "ideal_size_output";
2023-09-05 03:23:16 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Blur Image
* @description Blurs an image
*/
ImageBlurInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to blur */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Radius
* @description The blur radius
* @default 8
*/
radius?: number;
/**
2023-08-22 05:15:00 +00:00
* Blur Type
* @description The type of blur
* @default gaussian
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
blur_type?: "gaussian" | "box";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_blur
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_blur";
};
/**
2023-08-22 05:15:00 +00:00
* ImageCategory
* @description The category of an image.
2023-08-22 05:15:00 +00:00
*
* - GENERAL: The image is an output, init image, or otherwise an image without a specialized purpose.
* - MASK: The image is a mask image.
* - CONTROL: The image is a ControlNet control image.
* - USER: The image is a user-provide image.
2023-08-22 05:15:00 +00:00
* - OTHER: The image is some other type of image with a specialized purpose. To be used by external nodes.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
ImageCategory: "general" | "mask" | "control" | "user" | "other";
/**
2023-08-22 05:15:00 +00:00
* Extract Image Channel
* @description Gets a channel from an image.
*/
ImageChannelInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to get the channel from */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Channel
* @description The channel to get
* @default A
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
channel?: "A" | "R" | "G" | "B";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_chan
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_chan";
};
/**
* Multiply Image Channel
* @description Scale a specific color channel of an image.
*/
ImageChannelMultiplyInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to adjust */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* Channel
* @description Which channel to adjust
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
channel?: "Red (RGBA)" | "Green (RGBA)" | "Blue (RGBA)" | "Alpha (RGBA)" | "Cyan (CMYK)" | "Magenta (CMYK)" | "Yellow (CMYK)" | "Black (CMYK)" | "Hue (HSV)" | "Saturation (HSV)" | "Value (HSV)" | "Luminosity (LAB)" | "A (LAB)" | "B (LAB)" | "Y (YCbCr)" | "Cb (YCbCr)" | "Cr (YCbCr)";
/**
* Scale
* @description The amount to scale the channel by.
* @default 1
*/
scale?: number;
/**
* Invert Channel
* @description Invert the channel after scaling
* @default false
*/
invert_channel?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_channel_multiply
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_channel_multiply";
};
/**
* Offset Image Channel
* @description Add or subtract a value from a specific color channel of an image.
*/
ImageChannelOffsetInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to adjust */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* Channel
* @description Which channel to adjust
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
channel?: "Red (RGBA)" | "Green (RGBA)" | "Blue (RGBA)" | "Alpha (RGBA)" | "Cyan (CMYK)" | "Magenta (CMYK)" | "Yellow (CMYK)" | "Black (CMYK)" | "Hue (HSV)" | "Saturation (HSV)" | "Value (HSV)" | "Luminosity (LAB)" | "A (LAB)" | "B (LAB)" | "Y (YCbCr)" | "Cb (YCbCr)" | "Cr (YCbCr)";
/**
* Offset
* @description The amount to adjust the channel by
* @default 0
*/
offset?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_channel_offset
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_channel_offset";
};
/**
* Image Collection Primitive
* @description A collection of image primitive values
*/
ImageCollectionInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Collection
* @description The collection of image values
*/
2024-03-05 11:33:01 +00:00
collection?: components["schemas"]["ImageField"][];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default image_collection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "image_collection";
2024-02-09 20:14:15 +00:00
};
/**
* ImageCollectionOutput
* @description Base class for nodes that output a collection of images
*/
ImageCollectionOutput: {
/**
* Collection
* @description The output images
*/
2024-03-05 11:33:01 +00:00
collection: components["schemas"]["ImageField"][];
2024-02-09 20:14:15 +00:00
/**
* type
* @default image_collection_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "image_collection_output";
2024-02-09 20:14:15 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Convert Image Mode
* @description Converts an image to a different mode.
*/
ImageConvertInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to convert */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Mode
* @description The mode to convert to
* @default L
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
mode?: "L" | "RGB" | "RGBA" | "CMYK" | "YCbCr" | "LAB" | "HSV" | "I" | "F";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_conv
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_conv";
};
/**
2023-08-22 05:15:00 +00:00
* Crop Image
* @description Crops an image to a specified box. The box can be outside of the image.
*/
ImageCropInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to crop */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* X
* @description The left x coordinate of the crop rectangle
* @default 0
*/
x?: number;
/**
2023-08-22 05:15:00 +00:00
* Y
* @description The top y coordinate of the crop rectangle
* @default 0
*/
y?: number;
/**
2023-08-22 05:15:00 +00:00
* Width
* @description The width of the crop rectangle
* @default 512
*/
width?: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description The height of the crop rectangle
* @default 512
*/
height?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_crop
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_crop";
};
/**
2023-08-22 05:15:00 +00:00
* ImageDTO
* @description Deserialized image record, enriched for the frontend.
*/
ImageDTO: {
/**
2023-08-22 05:15:00 +00:00
* Image Name
* @description The unique name of the image.
*/
image_name: string;
/**
2023-08-22 05:15:00 +00:00
* Image Url
* @description The URL of the image.
*/
image_url: string;
/**
2023-08-22 05:15:00 +00:00
* Thumbnail Url
* @description The URL of the image's thumbnail.
*/
thumbnail_url: string;
/** @description The type of the image. */
2024-03-05 11:33:01 +00:00
image_origin: components["schemas"]["ResourceOrigin"];
/** @description The category of the image. */
2024-03-05 11:33:01 +00:00
image_category: components["schemas"]["ImageCategory"];
/**
2023-08-22 05:15:00 +00:00
* Width
* @description The width of the image in px.
*/
width: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description The height of the image in px.
*/
height: number;
/**
2023-08-22 05:15:00 +00:00
* Created At
* @description The created timestamp of the image.
*/
created_at: string;
/**
2023-08-22 05:15:00 +00:00
* Updated At
* @description The updated timestamp of the image.
*/
updated_at: string;
/**
2023-08-22 05:15:00 +00:00
* Deleted At
* @description The deleted timestamp of the image.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
deleted_at?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether this is an intermediate image.
*/
is_intermediate: boolean;
/**
2023-08-22 05:15:00 +00:00
* Session Id
* @description The session ID that generated this image, if it is a generated image.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
session_id?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Node Id
* @description The node ID that generated this image, if it is a generated image.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
node_id?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Starred
* @description Whether this image is starred.
*/
starred: boolean;
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Has Workflow
* @description Whether this image has a workflow.
*/
has_workflow: boolean;
/**
2023-08-22 05:15:00 +00:00
* Board Id
* @description The id of the board the image belongs to, if one exists.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
board_id?: string | null;
};
2024-02-09 20:14:15 +00:00
/**
2024-02-13 04:02:30 +00:00
* ImageField
* @description An image primitive field
*/
ImageField: {
/**
* Image Name
* @description The name of the image
*/
image_name: string;
};
/**
* Adjust Image Hue
* @description Adjusts the Hue of an image.
2024-02-09 20:14:15 +00:00
*/
2024-02-13 04:02:30 +00:00
ImageHueAdjustmentInvocation: {
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
2024-02-09 20:14:15 +00:00
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2024-02-09 20:14:15 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
2024-02-13 04:02:30 +00:00
/** @description The image to adjust */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2024-02-09 20:14:15 +00:00
/**
2024-02-13 04:02:30 +00:00
* Hue
* @description The degrees by which to rotate the hue, 0-360
* @default 0
2024-02-09 20:14:15 +00:00
*/
2024-02-13 04:02:30 +00:00
hue?: number;
2024-02-09 20:14:15 +00:00
/**
* type
2024-02-13 04:02:30 +00:00
* @default img_hue_adjust
2024-02-09 20:14:15 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_hue_adjust";
2024-02-09 20:14:15 +00:00
};
/**
2024-02-13 04:02:30 +00:00
* Inverse Lerp Image
* @description Inverse linear interpolation of all pixels of an image
*/
ImageInverseLerpInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to lerp */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Min
* @description The minimum input value
* @default 0
*/
min?: number;
/**
2023-08-22 05:15:00 +00:00
* Max
* @description The maximum input value
* @default 255
*/
max?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_ilerp
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_ilerp";
};
/**
2023-08-22 05:15:00 +00:00
* Image Primitive
* @description An image primitive value
*/
ImageInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to load */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default image
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "image";
};
/**
2023-08-22 05:15:00 +00:00
* Lerp Image
* @description Linear interpolation of all pixels of an image
*/
ImageLerpInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to lerp */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Min
* @description The minimum output value
* @default 0
*/
min?: number;
/**
2023-08-22 05:15:00 +00:00
* Max
* @description The maximum output value
* @default 255
*/
max?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_lerp
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_lerp";
};
/**
2023-08-22 05:15:00 +00:00
* Multiply Images
* @description Multiplies two images together using `PIL.ImageChops.multiply()`.
*/
ImageMultiplyInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The first image to multiply */
2024-03-05 11:33:01 +00:00
image1?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The second image to multiply */
2024-03-05 11:33:01 +00:00
image2?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_mul
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_mul";
};
2023-07-23 20:24:34 +00:00
/**
2023-08-22 05:15:00 +00:00
* Blur NSFW Image
2023-07-23 20:24:34 +00:00
* @description Add blur to NSFW-flagged images
*/
2023-07-24 12:25:39 +00:00
ImageNSFWBlurInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2023-07-23 20:24:34 +00:00
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-07-23 20:24:34 +00:00
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-07-23 20:24:34 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to check */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2023-07-23 20:24:34 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default img_nsfw
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-07-23 20:24:34 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "img_nsfw";
2023-07-23 20:24:34 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* ImageOutput
* @description Base class for nodes that output a single image
*/
ImageOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The output image */
2024-03-05 11:33:01 +00:00
image: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Width
* @description The width of the image in pixels
*/
width: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description The height of the image in pixels
*/
height: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default image_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "image_output";
};
/**
2023-12-06 12:33:54 +00:00
* Paste Image
* @description Pastes an image into another image.
*/
ImagePasteInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The base image */
2024-03-05 11:33:01 +00:00
base_image?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to paste */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The mask to use when pasting */
2024-03-05 11:33:01 +00:00
mask?: components["schemas"]["ImageField"] | null;
/**
2023-08-22 05:15:00 +00:00
* X
* @description The left x coordinate at which to paste the image
* @default 0
*/
x?: number;
/**
2023-08-22 05:15:00 +00:00
* Y
* @description The top y coordinate at which to paste the image
* @default 0
*/
y?: number;
/**
* Crop
* @description Crop to base image dimensions
* @default false
*/
crop?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-12-06 12:33:54 +00:00
* @default img_paste
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_paste";
};
/**
2023-08-22 05:15:00 +00:00
* ImageRecordChanges
* @description A set of changes to apply to an image record.
2023-08-22 05:15:00 +00:00
*
* Only limited changes are valid:
* - `image_category`: change the category of an image
* - `session_id`: change the session associated with an image
* - `is_intermediate`: change the image's `is_intermediate` flag
* - `starred`: change whether the image is starred
*/
ImageRecordChanges: {
/** @description The image's new category. */
2024-03-05 11:33:01 +00:00
image_category?: components["schemas"]["ImageCategory"] | null;
/**
2023-08-22 05:15:00 +00:00
* Session Id
* @description The image's new session ID.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
session_id?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description The image's new `is_intermediate` flag.
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
is_intermediate?: boolean | null;
/**
2023-08-22 05:15:00 +00:00
* Starred
* @description The image's new `starred` state
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
starred?: boolean | null;
[key: string]: unknown;
};
/**
2023-08-22 05:15:00 +00:00
* Resize Image
* @description Resizes an image to specific dimensions
*/
ImageResizeInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to resize */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Width
* @description The width to resize to (px)
* @default 512
*/
width?: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description The height to resize to (px)
* @default 512
*/
height?: number;
/**
2023-08-22 05:15:00 +00:00
* Resample Mode
* @description The resampling mode
* @default bicubic
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
resample_mode?: "nearest" | "box" | "bilinear" | "hamming" | "bicubic" | "lanczos";
2023-08-29 00:20:55 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_resize
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_resize";
};
/**
2023-08-22 05:15:00 +00:00
* Scale Image
* @description Scales an image by a factor
*/
ImageScaleInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to scale */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Scale Factor
* @description The factor by which to scale the image
2023-07-17 11:08:53 +00:00
* @default 2
*/
2023-07-17 11:08:53 +00:00
scale_factor?: number;
/**
2023-08-22 05:15:00 +00:00
* Resample Mode
* @description The resampling mode
* @default bicubic
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
resample_mode?: "nearest" | "box" | "bilinear" | "hamming" | "bicubic" | "lanczos";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_scale
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_scale";
};
/**
2023-08-22 05:15:00 +00:00
* Image to Latents
* @description Encodes an image into latents.
*/
ImageToLatentsInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to encode */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description VAE */
2024-03-05 11:33:01 +00:00
vae?: components["schemas"]["VaeField"];
/**
2023-08-22 05:15:00 +00:00
* Tiled
* @description Processing using overlapping tiles (reduce memory consumption)
* @default false
*/
tiled?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Fp32
* @description Whether or not to use full float32 precision
2024-01-24 13:15:54 +00:00
* @default false
*/
fp32?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default i2l
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "i2l";
};
/**
2023-08-22 05:15:00 +00:00
* ImageUrlsDTO
* @description The URLs for an image and its thumbnail.
*/
ImageUrlsDTO: {
/**
2023-08-22 05:15:00 +00:00
* Image Name
* @description The unique name of the image.
*/
image_name: string;
/**
2023-08-22 05:15:00 +00:00
* Image Url
* @description The URL of the image.
*/
image_url: string;
/**
2023-08-22 05:15:00 +00:00
* Thumbnail Url
* @description The URL of the image's thumbnail.
*/
thumbnail_url: string;
};
2023-07-24 12:25:39 +00:00
/**
2023-08-22 05:15:00 +00:00
* Add Invisible Watermark
2023-07-24 12:25:39 +00:00
* @description Add an invisible watermark to an image
*/
ImageWatermarkInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2023-07-24 12:25:39 +00:00
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-07-24 12:25:39 +00:00
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-07-24 12:25:39 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to check */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2023-07-24 12:25:39 +00:00
/**
2023-08-22 05:15:00 +00:00
* Text
* @description Watermark text
2023-07-24 12:25:39 +00:00
* @default InvokeAI
*/
text?: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default img_watermark
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "img_watermark";
2023-07-24 12:25:39 +00:00
};
/** ImagesDownloaded */
ImagesDownloaded: {
/**
* Response
2024-02-20 11:15:14 +00:00
* @description The message to display to the user when images begin downloading
*/
response?: string | null;
/**
* Bulk Download Item Name
* @description The name of the bulk download item for which events will be emitted
*/
2024-02-20 11:15:14 +00:00
bulk_download_item_name?: string | null;
};
/** ImagesUpdatedFromListResult */
ImagesUpdatedFromListResult: {
/**
2023-08-22 05:15:00 +00:00
* Updated Image Names
* @description The image names that were updated
*/
2023-08-22 05:15:00 +00:00
updated_image_names: string[];
};
/**
2023-08-22 05:15:00 +00:00
* Solid Color Infill
* @description Infills transparent areas of an image with a solid color
*/
InfillColorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to infill */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* @description The color to use to infill
* @default {
2023-11-16 00:21:28 +00:00
* "r": 127,
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* "g": 127,
2023-11-16 00:21:28 +00:00
* "b": 127,
* "a": 255
* }
*/
2024-03-05 11:33:01 +00:00
color?: components["schemas"]["ColorField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default infill_rgba
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "infill_rgba";
};
/**
2023-08-22 05:15:00 +00:00
* PatchMatch Infill
* @description Infills transparent areas of an image using the PatchMatch algorithm
*/
InfillPatchMatchInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to infill */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2023-09-05 02:17:23 +00:00
/**
* Downscale
* @description Run patchmatch on downscaled image to speedup infill
* @default 2
*/
downscale?: number;
/**
* Resample Mode
* @description The resampling mode
* @default bicubic
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
resample_mode?: "nearest" | "box" | "bilinear" | "hamming" | "bicubic" | "lanczos";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default infill_patchmatch
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "infill_patchmatch";
};
/**
2023-08-22 05:15:00 +00:00
* Tile Infill
* @description Infills transparent areas of an image with tiles of the image
*/
InfillTileInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to infill */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Tile Size
* @description The tile size (px)
* @default 32
*/
tile_size?: number;
/**
2023-08-22 05:15:00 +00:00
* Seed
* @description The seed to use for tile generation (omit for random)
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default 0
*/
seed?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default infill_tile
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "infill_tile";
};
/**
* InstallStatus
* @description State of an install job running in the background.
* @enum {string}
*/
InstallStatus: "waiting" | "downloading" | "downloads_done" | "running" | "completed" | "error" | "cancelled";
/**
* Integer Collection Primitive
* @description A collection of integer primitive values
*/
IntegerCollectionInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Collection
* @description The collection of integer values
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default []
*/
2023-08-22 05:15:00 +00:00
collection?: number[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default integer_collection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "integer_collection";
};
/**
2023-08-22 05:15:00 +00:00
* IntegerCollectionOutput
* @description Base class for nodes that output a collection of integers
*/
IntegerCollectionOutput: {
/**
* Collection
* @description The int collection
*/
collection: number[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default integer_collection_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "integer_collection_output";
};
/**
2023-08-22 05:15:00 +00:00
* Integer Primitive
* @description An integer primitive value
*/
IntegerInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Value
* @description The integer value
* @default 0
*/
value?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default integer
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "integer";
};
/**
* Integer Math
* @description Performs integer math.
*/
IntegerMathInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Operation
* @description The operation to perform
* @default ADD
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
operation?: "ADD" | "SUB" | "MUL" | "DIV" | "EXP" | "MOD" | "ABS" | "MIN" | "MAX";
/**
* A
* @description The first number
2023-11-08 06:28:37 +00:00
* @default 1
*/
a?: number;
/**
* B
* @description The second number
2023-11-08 06:28:37 +00:00
* @default 1
*/
b?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default integer_math
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "integer_math";
};
/**
2023-08-22 05:15:00 +00:00
* IntegerOutput
* @description Base class for nodes that output a single integer
*/
IntegerOutput: {
/**
* Value
* @description The output integer
*/
value: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default integer_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "integer_output";
};
/** InvocationCacheStatus */
InvocationCacheStatus: {
/**
* Size
* @description The current size of the invocation cache
*/
size: number;
/**
* Hits
* @description The number of cache hits
*/
hits: number;
/**
* Misses
* @description The number of cache misses
*/
misses: number;
/**
* Enabled
* @description Whether the invocation cache is enabled
*/
enabled: boolean;
/**
* Max Size
* @description The maximum size of the invocation cache
*/
max_size: number;
};
/**
2023-08-22 05:15:00 +00:00
* IterateInvocation
* @description Iterates over a list of items
*/
IterateInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Collection
* @description The list of items to iterate over
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default []
*/
2023-08-22 05:15:00 +00:00
collection?: unknown[];
/**
2023-08-22 05:15:00 +00:00
* Index
* @description The index, will be provided on executed iterators
* @default 0
*/
index?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default iterate
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "iterate";
};
/**
2023-08-22 05:15:00 +00:00
* IterateInvocationOutput
* @description Used to connect iteration outputs. Will be expanded to a specific output.
*/
IterateInvocationOutput: {
/**
* Collection Item
* @description The item being iterated over
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
item: unknown;
2023-12-06 12:33:54 +00:00
/**
* Index
* @description The index of the item
*/
index: number;
/**
* Total
* @description The total number of items
*/
total: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default iterate_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "iterate_output";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
JsonValue: unknown;
2023-08-23 19:25:24 +00:00
/**
* LaMa Infill
* @description Infills transparent areas of an image using the LaMa model
*/
LaMaInfillInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2023-08-23 19:25:24 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-08-23 19:25:24 +00:00
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-08-23 19:25:24 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to infill */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2023-08-23 19:25:24 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-23 19:25:24 +00:00
* @default infill_lama
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-08-23 19:25:24 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "infill_lama";
2023-08-23 19:25:24 +00:00
};
/**
* Latents Collection Primitive
* @description A collection of latents tensor primitive values
*/
LatentsCollectionInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Collection
* @description The collection of latents tensors
*/
2024-03-05 11:33:01 +00:00
collection?: components["schemas"]["LatentsField"][];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default latents_collection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "latents_collection";
};
/**
2023-08-22 05:15:00 +00:00
* LatentsCollectionOutput
* @description Base class for nodes that output a collection of latents tensors
*/
LatentsCollectionOutput: {
/**
* Collection
* @description Latents tensor
*/
2024-03-05 11:33:01 +00:00
collection: components["schemas"]["LatentsField"][];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default latents_collection_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "latents_collection_output";
};
/**
2023-08-22 05:15:00 +00:00
* LatentsField
* @description A latents tensor primitive field
*/
LatentsField: {
/**
2023-08-22 05:15:00 +00:00
* Latents Name
* @description The name of the latents
*/
latents_name: string;
/**
2023-08-22 05:15:00 +00:00
* Seed
* @description Seed used to generate this latents
2024-02-17 09:02:51 +00:00
* @default null
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
seed?: number | null;
};
/**
2023-08-22 05:15:00 +00:00
* Latents Primitive
* @description A latents tensor primitive value
*/
LatentsInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The latents tensor */
2024-03-05 11:33:01 +00:00
latents?: components["schemas"]["LatentsField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default latents
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "latents";
};
/**
2023-08-22 05:15:00 +00:00
* LatentsOutput
* @description Base class for nodes that output a single latents tensor
*/
LatentsOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents: components["schemas"]["LatentsField"];
/**
2023-08-22 05:15:00 +00:00
* Width
* @description Width of output (px)
*/
width: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description Height of output (px)
*/
height: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default latents_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "latents_output";
};
/**
2023-08-22 05:15:00 +00:00
* Latents to Image
* @description Generates an image from latents.
*/
LatentsToImageInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents?: components["schemas"]["LatentsField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description VAE */
2024-03-05 11:33:01 +00:00
vae?: components["schemas"]["VaeField"];
/**
2023-08-22 05:15:00 +00:00
* Tiled
* @description Processing using overlapping tiles (reduce memory consumption)
* @default false
*/
tiled?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Fp32
* @description Whether or not to use full float32 precision
2024-01-24 13:15:54 +00:00
* @default false
*/
fp32?: boolean;
2023-07-12 15:15:09 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default l2i
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "l2i";
};
/**
2023-08-22 05:15:00 +00:00
* Leres (Depth) Processor
* @description Applies leres processing to image
*/
LeresImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Thr A
* @description Leres parameter `thr_a`
* @default 0
*/
thr_a?: number;
/**
2023-08-22 05:15:00 +00:00
* Thr B
* @description Leres parameter `thr_b`
* @default 0
*/
thr_b?: number;
/**
2023-08-22 05:15:00 +00:00
* Boost
* @description Whether to use boost mode
* @default false
*/
boost?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default leres_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "leres_image_processor";
};
/**
2023-08-22 05:15:00 +00:00
* Lineart Anime Processor
* @description Applies line art anime processing to image
*/
LineartAnimeImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default lineart_anime_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lineart_anime_image_processor";
};
/**
2023-08-22 05:15:00 +00:00
* Lineart Processor
* @description Applies line art processing to image
*/
LineartImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Coarse
* @description Whether to use coarse mode
* @default false
*/
coarse?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default lineart_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lineart_image_processor";
};
2023-11-16 00:21:28 +00:00
/**
* LoRADiffusersConfig
* @description Model config for LoRA/Diffusers models.
2023-11-16 00:21:28 +00:00
*/
LoRADiffusersConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
/**
* Description
* @description Model description
*/
description?: string | null;
/**
* Source
* @description The original source of the model (path, URL or repo_id).
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
*/
source_api_response?: string | null;
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2023-11-16 00:21:28 +00:00
/**
* Type
* @default lora
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lora";
2023-11-16 00:21:28 +00:00
/**
* Format
* @default diffusers
* @constant
2023-11-16 00:21:28 +00:00
*/
format: "diffusers";
};
2024-03-06 08:37:52 +00:00
/** LoRAField */
LoRAField: {
/** @description Info to load lora model */
lora: components["schemas"]["ModelField"];
/**
* Weight
* @description Weight to apply to lora model
*/
weight: number;
};
/**
* LoRALyCORISConfig
* @description Model config for LoRA/Lycoris models.
*/
LoRALyCORISConfig: {
2023-11-16 00:21:28 +00:00
/**
* Key
* @description A unique key for this model.
2023-11-16 00:21:28 +00:00
*/
key: string;
2023-11-16 00:21:28 +00:00
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2023-11-16 00:21:28 +00:00
*/
path: string;
2023-11-16 00:21:28 +00:00
/**
* Name
* @description Name of the model.
2023-11-16 00:21:28 +00:00
*/
name: string;
/** @description The base model. */
base: components["schemas"]["BaseModelType"];
2024-02-15 11:15:21 +00:00
/**
* Description
* @description Model description
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
description?: string | null;
/**
* Source
* @description The original source of the model (path, URL or repo_id).
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
*/
source_api_response?: string | null;
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
/**
* Type
* @default lora
* @constant
2023-11-16 00:21:28 +00:00
*/
type: "lora";
2024-02-15 11:15:21 +00:00
/**
* Format
* @default lycoris
* @constant
2024-02-15 11:15:21 +00:00
*/
format: "lycoris";
2023-11-16 00:21:28 +00:00
};
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* LoRAMetadataField
* @description LoRA Metadata Field
2023-07-12 15:15:09 +00:00
*/
LoRAMetadataField: {
/** @description LoRA model to load */
2024-03-06 08:37:52 +00:00
model: components["schemas"]["ModelMetadataField"];
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Weight
* @description The weight at which the LoRA is applied to each model
2023-07-12 15:15:09 +00:00
*/
weight: number;
};
/**
* LocalModelSource
* @description A local file or directory path.
*/
LocalModelSource: {
/** Path */
path: string;
/**
* Inplace
* @default false
*/
inplace?: boolean | null;
/**
* Type
* @default local
* @constant
*/
2024-03-05 11:33:01 +00:00
type?: "local";
};
2023-07-20 15:45:54 +00:00
/**
2023-08-22 05:15:00 +00:00
* LogLevel
2023-07-20 15:45:54 +00:00
* @enum {integer}
*/
LogLevel: 0 | 10 | 20 | 30 | 40 | 50;
/**
2023-08-22 05:15:00 +00:00
* LoRA
* @description Apply selected lora to unet and text_encoder.
*/
LoraLoaderInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* LoRA
* @description LoRA model to load
*/
2024-03-06 08:37:52 +00:00
lora: components["schemas"]["ModelField"];
/**
2023-08-22 05:15:00 +00:00
* Weight
* @description The weight at which the LoRA is applied to each model
* @default 0.75
*/
weight?: number;
/**
2023-08-22 05:15:00 +00:00
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet?: components["schemas"]["UNetField"] | null;
/**
2023-08-22 05:15:00 +00:00
* CLIP
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip?: components["schemas"]["ClipField"] | null;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default lora_loader
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lora_loader";
};
/**
2023-08-22 05:15:00 +00:00
* LoraLoaderOutput
* @description Model loader output
*/
LoraLoaderOutput: {
/**
2023-08-22 05:15:00 +00:00
* UNet
* @description UNet (scheduler, LoRAs)
2024-02-17 09:02:51 +00:00
* @default null
*/
2024-03-05 11:33:01 +00:00
unet: components["schemas"]["UNetField"] | null;
/**
2023-08-22 05:15:00 +00:00
* CLIP
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
2024-02-17 09:02:51 +00:00
* @default null
*/
2024-03-05 11:33:01 +00:00
clip: components["schemas"]["ClipField"] | null;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default lora_loader_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lora_loader_output";
};
2023-11-16 00:21:28 +00:00
/**
* MainCheckpointConfig
* @description Model config for main checkpoint models.
*/
MainCheckpointConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
2023-11-16 00:21:28 +00:00
/**
* Description
* @description Model description
2023-11-16 00:21:28 +00:00
*/
description?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
2023-11-16 00:21:28 +00:00
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
2023-11-16 00:21:28 +00:00
*/
source_api_response?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
2023-11-16 00:21:28 +00:00
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2023-11-16 00:21:28 +00:00
/**
* Format
* @default checkpoint
* @constant
2023-11-16 00:21:28 +00:00
*/
2024-03-05 11:33:01 +00:00
format: "checkpoint";
2024-02-15 11:15:21 +00:00
/**
* Config Path
* @description path to the checkpoint model config file
2024-02-15 11:15:21 +00:00
*/
config_path: string;
2023-11-16 00:21:28 +00:00
/**
* Converted At
* @description When this model was last converted to diffusers
2023-11-16 00:21:28 +00:00
*/
converted_at?: number | null;
2024-02-15 11:15:21 +00:00
/**
* Type
* @default main
* @constant
2024-02-15 11:15:21 +00:00
*/
type: "main";
2023-11-16 00:21:28 +00:00
/** @default normal */
2024-03-05 11:33:01 +00:00
variant?: components["schemas"]["ModelVariantType"];
2024-02-15 11:15:21 +00:00
/** @default epsilon */
2024-03-05 11:33:01 +00:00
prediction_type?: components["schemas"]["SchedulerPredictionType"];
2024-02-15 11:15:21 +00:00
/**
* Upcast Attention
* @default false
*/
upcast_attention?: boolean;
2023-11-16 00:21:28 +00:00
};
/**
* MainDiffusersConfig
* @description Model config for main diffusers models.
*/
MainDiffusersConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
/**
* Description
* @description Model description
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
description?: string | null;
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
2024-02-15 11:15:21 +00:00
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
2024-02-15 11:15:21 +00:00
*/
source_api_response?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
2023-11-16 00:21:28 +00:00
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2023-11-16 00:21:28 +00:00
/**
* Format
* @default diffusers
* @constant
2023-11-16 00:21:28 +00:00
*/
format: "diffusers";
2024-02-15 11:15:21 +00:00
/** @default */
repo_variant?: components["schemas"]["ModelRepoVariant"] | null;
/**
* Type
* @default main
* @constant
*/
type: "main";
2023-11-16 00:21:28 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Main Model
* @description Loads a main model, outputting its submodels.
*/
MainModelLoaderInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Main model (UNet, VAE, CLIP) to load */
2024-03-06 08:37:52 +00:00
model: components["schemas"]["ModelField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default main_model_loader
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "main_model_loader";
};
2023-08-11 18:15:59 +00:00
/**
* Combine Masks
2023-08-11 18:15:59 +00:00
* @description Combine two masks together by multiplying them using `PIL.ImageChops.multiply()`.
*/
MaskCombineInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2023-08-11 18:15:59 +00:00
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-08-11 18:15:59 +00:00
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-08-11 18:15:59 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The first mask to combine */
2024-03-05 11:33:01 +00:00
mask1?: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The second image to combine */
2024-03-05 11:33:01 +00:00
mask2?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default mask_combine
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "mask_combine";
2023-08-11 18:15:59 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Mask Edge
* @description Applies an edge mask to an image
*/
MaskEdgeInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to apply the mask to */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Edge Size
* @description The size of the edge
*/
edge_size?: number;
/**
2023-08-22 05:15:00 +00:00
* Edge Blur
* @description The amount of blur on the edge
*/
edge_blur?: number;
/**
2023-08-22 05:15:00 +00:00
* Low Threshold
* @description First threshold for the hysteresis procedure in Canny edge detection
*/
low_threshold?: number;
/**
2023-08-22 05:15:00 +00:00
* High Threshold
* @description Second threshold for the hysteresis procedure in Canny edge detection
*/
2024-02-09 20:14:15 +00:00
high_threshold?: number;
/**
* type
* @default mask_edge
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "mask_edge";
2024-02-09 20:14:15 +00:00
};
/**
* Mask from Alpha
* @description Extracts the alpha channel of an image as a mask.
*/
MaskFromAlphaInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
2024-02-09 20:14:15 +00:00
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
2024-02-09 20:14:15 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The image to create the mask from */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
2024-02-09 20:14:15 +00:00
/**
* Invert
* @description Whether or not to invert the mask
* @default false
*/
invert?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2024-02-09 20:14:15 +00:00
* @default tomask
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "tomask";
};
/**
2023-08-22 05:15:00 +00:00
* Mediapipe Face Processor
* @description Applies mediapipe face processing to image
*/
MediapipeFaceProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Max Faces
* @description Maximum number of faces to detect
* @default 1
*/
max_faces?: number;
/**
2023-08-22 05:15:00 +00:00
* Min Confidence
* @description Minimum confidence for face detection
* @default 0.5
*/
min_confidence?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default mediapipe_face_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "mediapipe_face_processor";
};
/**
* Metadata Merge
* @description Merged a collection of MetadataDict into a single MetadataDict.
*/
MergeMetadataInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Collection
* @description Collection of Metadata
*/
2024-03-05 11:33:01 +00:00
collection?: components["schemas"]["MetadataField"][];
/**
* type
* @default merge_metadata
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "merge_metadata";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Merge Tiles to Image
* @description Merge multiple tile images into a single image.
*/
MergeTilesToImageInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Tiles With Images
* @description A list of tile images with tile properties.
*/
2024-03-05 11:33:01 +00:00
tiles_with_images?: components["schemas"]["TileWithImage"][];
/**
* Blend Mode
* @description blending type Linear or Seam
* @default Seam
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
blend_mode?: "Linear" | "Seam";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Blend Amount
* @description The amount to blend adjacent tiles in pixels. Must be <= the amount of overlap between adjacent tiles.
* @default 32
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
*/
blend_amount?: number;
/**
* type
* @default merge_tiles_to_image
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "merge_tiles_to_image";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
2023-07-12 15:15:09 +00:00
/**
* MetadataField
* @description Pydantic model for metadata with custom root of type dict[str, Any].
* Metadata is stored without a strict schema.
*/
2024-03-06 08:37:52 +00:00
MetadataField: Record<string, unknown>;
/**
* Metadata
* @description Takes a MetadataItem or collection of MetadataItems and outputs a MetadataDict.
2023-07-12 15:15:09 +00:00
*/
MetadataInvocation: {
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-07-12 15:15:09 +00:00
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-07-12 15:15:09 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
2023-07-12 15:15:09 +00:00
/**
* Items
* @description A single metadata item or collection of metadata items
*/
2024-03-05 11:33:01 +00:00
items?: components["schemas"]["MetadataItemField"][] | components["schemas"]["MetadataItemField"];
/**
* type
* @default metadata
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "metadata";
};
/** MetadataItemField */
MetadataItemField: {
/**
* Label
* @description Label for this metadata item
*/
label: string;
/**
* Value
* @description The value for this metadata item (may be any type)
*/
value: unknown;
};
/**
* Metadata Item
* @description Used to create an arbitrary metadata item. Provide "label" and make a connection to "value" to store that data as the value.
*/
MetadataItemInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Label
* @description Label for this metadata item
*/
label?: string;
/**
* Value
* @description The value for this metadata item (may be any type)
*/
value?: unknown;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default metadata_item
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "metadata_item";
2023-07-12 15:15:09 +00:00
};
/**
* MetadataItemOutput
* @description Metadata Item Output
2023-07-12 15:15:09 +00:00
*/
MetadataItemOutput: {
/** @description Metadata Item */
2024-03-05 11:33:01 +00:00
item: components["schemas"]["MetadataItemField"];
2023-07-12 15:15:09 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default metadata_item_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-07-12 15:15:09 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "metadata_item_output";
};
/** MetadataOutput */
MetadataOutput: {
/** @description Metadata Dict */
2024-03-05 11:33:01 +00:00
metadata: components["schemas"]["MetadataField"];
/**
* type
* @default metadata_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "metadata_output";
2023-07-12 15:15:09 +00:00
};
/**
* Midas Depth Processor
* @description Applies Midas depth processing to image
*/
MidasDepthImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* A Mult
* @description Midas parameter `a_mult` (a = a_mult * PI)
* @default 2
*/
a_mult?: number;
/**
2023-08-22 05:15:00 +00:00
* Bg Th
* @description Midas parameter `bg_th`
* @default 0.1
*/
bg_th?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default midas_depth_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "midas_depth_image_processor";
};
/**
2023-08-22 05:15:00 +00:00
* MLSD Processor
* @description Applies MLSD processing to image
*/
MlsdImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Thr V
* @description MLSD parameter `thr_v`
* @default 0.1
*/
thr_v?: number;
/**
2023-08-22 05:15:00 +00:00
* Thr D
* @description MLSD parameter `thr_d`
* @default 0.1
*/
thr_d?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default mlsd_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "mlsd_image_processor";
};
/** ModelDefaultSettings */
ModelDefaultSettings: {
/** Vae */
vae: string | null;
/** Vae Precision */
vae_precision: string | null;
/** Scheduler */
scheduler: ("ddim" | "ddpm" | "deis" | "lms" | "lms_k" | "pndm" | "heun" | "heun_k" | "euler" | "euler_k" | "euler_a" | "kdpm_2" | "kdpm_2_a" | "dpmpp_2s" | "dpmpp_2s_k" | "dpmpp_2m" | "dpmpp_2m_k" | "dpmpp_2m_sde" | "dpmpp_2m_sde_k" | "dpmpp_sde" | "dpmpp_sde_k" | "unipc" | "lcm") | null;
/** Steps */
steps: number | null;
/** Cfg Scale */
cfg_scale: number | null;
/** Cfg Rescale Multiplier */
cfg_rescale_multiplier: number | null;
};
2024-03-06 08:37:52 +00:00
/** ModelField */
ModelField: {
/**
2024-02-15 11:15:21 +00:00
* Key
2024-03-06 08:37:52 +00:00
* @description Key of the model
*/
2024-02-15 11:15:21 +00:00
key: string;
2024-02-17 09:02:51 +00:00
/**
2024-03-06 08:37:52 +00:00
* @description Submodel type
2024-02-17 09:02:51 +00:00
* @default null
*/
2024-03-05 11:33:01 +00:00
submodel_type?: components["schemas"]["SubModelType"] | null;
};
2024-03-06 08:37:52 +00:00
/**
* ModelFormat
* @description Storage format of model.
* @enum {string}
*/
ModelFormat: "diffusers" | "checkpoint" | "lycoris" | "onnx" | "olive" | "embedding_file" | "embedding_folder" | "invokeai";
/**
* ModelInstallJob
* @description Object that tracks the current status of an install request.
*/
ModelInstallJob: {
2024-01-22 22:28:24 +00:00
/**
* Id
* @description Unique ID for this job
*/
id: number;
/**
* @description Current status of install process
* @default waiting
*/
2024-03-05 11:33:01 +00:00
status?: components["schemas"]["InstallStatus"];
2024-02-23 15:41:58 +00:00
/**
* Error Reason
* @description Information about why the job failed
*/
error_reason?: string | null;
/**
* Config In
* @description Configuration information (e.g. 'description') to apply to model.
*/
config_in?: Record<string, never>;
/**
* Config Out
* @description After successful installation, this will hold the configuration object.
*/
config_out?: (components["schemas"]["MainDiffusersConfig"] | components["schemas"]["MainCheckpointConfig"] | components["schemas"]["VAEDiffusersConfig"] | components["schemas"]["VAECheckpointConfig"] | components["schemas"]["ControlNetDiffusersConfig"] | components["schemas"]["ControlNetCheckpointConfig"] | components["schemas"]["LoRALyCORISConfig"] | components["schemas"]["LoRADiffusersConfig"] | components["schemas"]["TextualInversionFileConfig"] | components["schemas"]["TextualInversionFolderConfig"] | components["schemas"]["IPAdapterConfig"] | components["schemas"]["T2IAdapterConfig"] | components["schemas"]["CLIPVisionDiffusersConfig"]) | null;
/**
* Inplace
* @description Leave model in its current location; otherwise install under models directory
* @default false
*/
inplace?: boolean;
/**
* Source
* @description Source (URL, repo_id, or local path) of model
*/
2024-03-05 11:33:01 +00:00
source: components["schemas"]["LocalModelSource"] | components["schemas"]["HFModelSource"] | components["schemas"]["CivitaiModelSource"] | components["schemas"]["URLModelSource"];
/**
* Local Path
* Format: path
* @description Path to locally-downloaded model; may be the same as the source
*/
local_path: string;
/**
2024-01-22 22:28:24 +00:00
* Bytes
* @description For a remote model, the number of bytes downloaded so far (may not be available)
2024-02-09 20:14:15 +00:00
* @default 0
*/
2024-02-09 20:14:15 +00:00
bytes?: number;
/**
2024-01-22 22:28:24 +00:00
* Total Bytes
* @description Total size of the model to be installed
* @default 0
*/
2024-01-22 22:28:24 +00:00
total_bytes?: number;
/**
* Source Metadata
* @description Metadata provided by the model source
*/
2024-03-05 11:33:01 +00:00
source_metadata?: (components["schemas"]["BaseMetadata"] | components["schemas"]["HuggingFaceMetadata"] | components["schemas"]["CivitaiMetadata"]) | null;
2024-01-22 22:28:24 +00:00
/**
* Download Parts
* @description Download jobs contributing to this install
*/
2024-03-05 11:33:01 +00:00
download_parts?: components["schemas"]["DownloadJob"][];
/**
* Error
* @description On an error condition, this field will contain the text of the exception
*/
error?: string | null;
/**
* Error Traceback
* @description On an error condition, this field will contain the exception traceback
*/
error_traceback?: string | null;
};
/**
2023-08-22 05:15:00 +00:00
* ModelLoaderOutput
* @description Model loader output
*/
ModelLoaderOutput: {
/**
2023-08-22 05:15:00 +00:00
* VAE
* @description VAE
*/
2024-03-05 11:33:01 +00:00
vae: components["schemas"]["VaeField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default model_loader_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "model_loader_output";
2023-11-08 06:28:37 +00:00
/**
* CLIP
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip: components["schemas"]["ClipField"];
2023-11-08 06:28:37 +00:00
/**
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet: components["schemas"]["UNetField"];
};
2024-03-06 08:37:52 +00:00
/**
* ModelMetadataField
* @description Model Metadata Field
*/
ModelMetadataField: {
/** Key */
key: string;
/** Hash */
hash: string;
/** Name */
name: string;
base: components["schemas"]["BaseModelType"];
type: components["schemas"]["ModelType"];
};
/**
* ModelRecordChanges
* @description A set of changes to apply to a model.
*/
ModelRecordChanges: {
/**
* Name
* @description Name of the model.
*/
name?: string | null;
2024-03-05 11:33:01 +00:00
/**
* Path
* @description Path to the model.
*/
path?: string | null;
/**
* Description
* @description Model description
*/
description?: string | null;
/** @description The base model. */
base?: components["schemas"]["BaseModelType"] | null;
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
/** @description The variant of the model. */
variant?: components["schemas"]["ModelVariantType"] | null;
/** @description The prediction type of the model. */
prediction_type?: components["schemas"]["SchedulerPredictionType"] | null;
/**
* Upcast Attention
* @description Whether to upcast attention.
*/
upcast_attention?: boolean | null;
/**
* Config Path
* @description Path to config file for model
*/
config_path?: string | null;
};
2024-01-22 22:28:24 +00:00
/**
* ModelRepoVariant
* @description Various hugging face variants on the diffusers format.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
ModelRepoVariant: "" | "fp16" | "fp32" | "onnx" | "openvino" | "flax";
2024-01-22 22:28:24 +00:00
/**
* ModelSourceType
* @description Model source type.
* @enum {string}
2024-01-22 22:28:24 +00:00
*/
ModelSourceType: "path" | "url" | "hf_repo_id" | "civitai";
2024-02-15 11:15:21 +00:00
/**
* ModelType
* @description Model type.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
ModelType: "onnx" | "main" | "vae" | "lora" | "controlnet" | "embedding" | "ip_adapter" | "clip_vision" | "t2i_adapter";
2024-02-15 11:15:21 +00:00
/**
* ModelVariantType
* @description Variant type.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
ModelVariantType: "normal" | "inpaint" | "depth";
2024-02-15 11:15:21 +00:00
/**
* ModelsList
* @description Return list of configs.
*/
ModelsList: {
/** Models */
models: (components["schemas"]["MainDiffusersConfig"] | components["schemas"]["MainCheckpointConfig"] | components["schemas"]["VAEDiffusersConfig"] | components["schemas"]["VAECheckpointConfig"] | components["schemas"]["ControlNetDiffusersConfig"] | components["schemas"]["ControlNetCheckpointConfig"] | components["schemas"]["LoRALyCORISConfig"] | components["schemas"]["LoRADiffusersConfig"] | components["schemas"]["TextualInversionFileConfig"] | components["schemas"]["TextualInversionFolderConfig"] | components["schemas"]["IPAdapterConfig"] | components["schemas"]["T2IAdapterConfig"] | components["schemas"]["CLIPVisionDiffusersConfig"])[];
2024-02-15 11:15:21 +00:00
};
2023-08-16 01:59:19 +00:00
/**
2023-08-22 05:15:00 +00:00
* Multiply Integers
* @description Multiplies two numbers
*/
MultiplyInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* A
* @description The first number
* @default 0
*/
a?: number;
/**
2023-08-22 05:15:00 +00:00
* B
* @description The second number
* @default 0
*/
b?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default mul
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "mul";
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** NodeFieldValue */
NodeFieldValue: {
/**
* Node Path
* @description The node into which this batch data item will be substituted.
*/
node_path: string;
/**
* Field Name
* @description The field into which this batch data item will be substituted.
*/
field_name: string;
/**
* Value
* @description The value to substitute into the node/field.
*/
value: string | number;
};
2024-02-09 20:14:15 +00:00
/**
2024-02-13 04:02:30 +00:00
* Noise
* @description Generates latent noise.
*/
NoiseInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Seed
* @description Seed for random number generation
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default 0
*/
seed?: number;
/**
2023-08-22 05:15:00 +00:00
* Width
* @description Width of output (px)
* @default 512
*/
width?: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description Height of output (px)
* @default 512
*/
height?: number;
2023-06-27 03:57:41 +00:00
/**
2023-08-22 05:15:00 +00:00
* Use Cpu
* @description Use CPU for noise generation (for reproducible results across platforms)
2023-06-27 03:57:41 +00:00
* @default true
*/
use_cpu?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default noise
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "noise";
};
/**
2023-08-22 05:15:00 +00:00
* NoiseOutput
* @description Invocation noise output
*/
NoiseOutput: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Noise tensor */
2024-03-05 11:33:01 +00:00
noise: components["schemas"]["LatentsField"];
/**
2023-08-22 05:15:00 +00:00
* Width
* @description Width of output (px)
*/
width: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description Height of output (px)
*/
height: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default noise_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "noise_output";
};
/**
2023-08-22 05:15:00 +00:00
* Normal BAE Processor
* @description Applies NormalBae processing to image
*/
NormalbaeImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default normalbae_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "normalbae_image_processor";
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** OffsetPaginatedResults[BoardDTO] */
OffsetPaginatedResults_BoardDTO_: {
/**
feat: refactor services folder/module structure Refactor services folder/module structure. **Motivation** While working on our services I've repeatedly encountered circular imports and a general lack of clarity regarding where to put things. The structure introduced goes a long way towards resolving those issues, setting us up for a clean structure going forward. **Services** Services are now in their own folder with a few files: - `services/{service_name}/__init__.py`: init as needed, mostly empty now - `services/{service_name}/{service_name}_base.py`: the base class for the service - `services/{service_name}/{service_name}_{impl_type}.py`: the default concrete implementation of the service - typically one of `sqlite`, `default`, or `memory` - `services/{service_name}/{service_name}_common.py`: any common items - models, exceptions, utilities, etc Though it's a bit verbose to have the service name both as the folder name and the prefix for files, I found it is _extremely_ confusing to have all of the base classes just be named `base.py`. So, at the cost of some verbosity when importing things, I've included the service name in the filename. There are some minor logic changes. For example, in `InvocationProcessor`, instead of assigning the model manager service to a variable to be used later in the file, the service is used directly via the `Invoker`. **Shared** Things that are used across disparate services are in `services/shared/`: - `default_graphs.py`: previously in `services/` - `graphs.py`: previously in `services/` - `paginatation`: generic pagination models used in a few services - `sqlite`: the `SqliteDatabase` class, other sqlite-specific things
2023-09-24 08:11:07 +00:00
* Limit
* @description Limit of items to get
*/
feat: refactor services folder/module structure Refactor services folder/module structure. **Motivation** While working on our services I've repeatedly encountered circular imports and a general lack of clarity regarding where to put things. The structure introduced goes a long way towards resolving those issues, setting us up for a clean structure going forward. **Services** Services are now in their own folder with a few files: - `services/{service_name}/__init__.py`: init as needed, mostly empty now - `services/{service_name}/{service_name}_base.py`: the base class for the service - `services/{service_name}/{service_name}_{impl_type}.py`: the default concrete implementation of the service - typically one of `sqlite`, `default`, or `memory` - `services/{service_name}/{service_name}_common.py`: any common items - models, exceptions, utilities, etc Though it's a bit verbose to have the service name both as the folder name and the prefix for files, I found it is _extremely_ confusing to have all of the base classes just be named `base.py`. So, at the cost of some verbosity when importing things, I've included the service name in the filename. There are some minor logic changes. For example, in `InvocationProcessor`, instead of assigning the model manager service to a variable to be used later in the file, the service is used directly via the `Invoker`. **Shared** Things that are used across disparate services are in `services/shared/`: - `default_graphs.py`: previously in `services/` - `graphs.py`: previously in `services/` - `paginatation`: generic pagination models used in a few services - `sqlite`: the `SqliteDatabase` class, other sqlite-specific things
2023-09-24 08:11:07 +00:00
limit: number;
/**
2023-08-22 05:15:00 +00:00
* Offset
* @description Offset from which to retrieve items
*/
offset: number;
/**
2023-08-22 05:15:00 +00:00
* Total
* @description Total number of items in result
*/
total: number;
feat: refactor services folder/module structure Refactor services folder/module structure. **Motivation** While working on our services I've repeatedly encountered circular imports and a general lack of clarity regarding where to put things. The structure introduced goes a long way towards resolving those issues, setting us up for a clean structure going forward. **Services** Services are now in their own folder with a few files: - `services/{service_name}/__init__.py`: init as needed, mostly empty now - `services/{service_name}/{service_name}_base.py`: the base class for the service - `services/{service_name}/{service_name}_{impl_type}.py`: the default concrete implementation of the service - typically one of `sqlite`, `default`, or `memory` - `services/{service_name}/{service_name}_common.py`: any common items - models, exceptions, utilities, etc Though it's a bit verbose to have the service name both as the folder name and the prefix for files, I found it is _extremely_ confusing to have all of the base classes just be named `base.py`. So, at the cost of some verbosity when importing things, I've included the service name in the filename. There are some minor logic changes. For example, in `InvocationProcessor`, instead of assigning the model manager service to a variable to be used later in the file, the service is used directly via the `Invoker`. **Shared** Things that are used across disparate services are in `services/shared/`: - `default_graphs.py`: previously in `services/` - `graphs.py`: previously in `services/` - `paginatation`: generic pagination models used in a few services - `sqlite`: the `SqliteDatabase` class, other sqlite-specific things
2023-09-24 08:11:07 +00:00
/**
* Items
* @description Items
*/
2024-03-05 11:33:01 +00:00
items: components["schemas"]["BoardDTO"][];
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** OffsetPaginatedResults[ImageDTO] */
OffsetPaginatedResults_ImageDTO_: {
/**
feat: refactor services folder/module structure Refactor services folder/module structure. **Motivation** While working on our services I've repeatedly encountered circular imports and a general lack of clarity regarding where to put things. The structure introduced goes a long way towards resolving those issues, setting us up for a clean structure going forward. **Services** Services are now in their own folder with a few files: - `services/{service_name}/__init__.py`: init as needed, mostly empty now - `services/{service_name}/{service_name}_base.py`: the base class for the service - `services/{service_name}/{service_name}_{impl_type}.py`: the default concrete implementation of the service - typically one of `sqlite`, `default`, or `memory` - `services/{service_name}/{service_name}_common.py`: any common items - models, exceptions, utilities, etc Though it's a bit verbose to have the service name both as the folder name and the prefix for files, I found it is _extremely_ confusing to have all of the base classes just be named `base.py`. So, at the cost of some verbosity when importing things, I've included the service name in the filename. There are some minor logic changes. For example, in `InvocationProcessor`, instead of assigning the model manager service to a variable to be used later in the file, the service is used directly via the `Invoker`. **Shared** Things that are used across disparate services are in `services/shared/`: - `default_graphs.py`: previously in `services/` - `graphs.py`: previously in `services/` - `paginatation`: generic pagination models used in a few services - `sqlite`: the `SqliteDatabase` class, other sqlite-specific things
2023-09-24 08:11:07 +00:00
* Limit
* @description Limit of items to get
*/
feat: refactor services folder/module structure Refactor services folder/module structure. **Motivation** While working on our services I've repeatedly encountered circular imports and a general lack of clarity regarding where to put things. The structure introduced goes a long way towards resolving those issues, setting us up for a clean structure going forward. **Services** Services are now in their own folder with a few files: - `services/{service_name}/__init__.py`: init as needed, mostly empty now - `services/{service_name}/{service_name}_base.py`: the base class for the service - `services/{service_name}/{service_name}_{impl_type}.py`: the default concrete implementation of the service - typically one of `sqlite`, `default`, or `memory` - `services/{service_name}/{service_name}_common.py`: any common items - models, exceptions, utilities, etc Though it's a bit verbose to have the service name both as the folder name and the prefix for files, I found it is _extremely_ confusing to have all of the base classes just be named `base.py`. So, at the cost of some verbosity when importing things, I've included the service name in the filename. There are some minor logic changes. For example, in `InvocationProcessor`, instead of assigning the model manager service to a variable to be used later in the file, the service is used directly via the `Invoker`. **Shared** Things that are used across disparate services are in `services/shared/`: - `default_graphs.py`: previously in `services/` - `graphs.py`: previously in `services/` - `paginatation`: generic pagination models used in a few services - `sqlite`: the `SqliteDatabase` class, other sqlite-specific things
2023-09-24 08:11:07 +00:00
limit: number;
/**
2023-08-22 05:15:00 +00:00
* Offset
* @description Offset from which to retrieve items
*/
offset: number;
/**
2023-08-22 05:15:00 +00:00
* Total
* @description Total number of items in result
*/
2024-02-13 04:02:30 +00:00
total: number;
/**
2024-02-13 04:02:30 +00:00
* Items
* @description Items
*/
2024-03-05 11:33:01 +00:00
items: components["schemas"]["ImageDTO"][];
};
/** PaginatedResults[WorkflowRecordListItemDTO] */
PaginatedResults_WorkflowRecordListItemDTO_: {
2024-01-22 22:28:24 +00:00
/**
* Page
* @description Current Page
*/
page: number;
/**
* Pages
* @description Total number of pages
*/
pages: number;
/**
* Per Page
* @description Number of items per page
*/
per_page: number;
/**
* Total
* @description Total number of items in result
*/
total: number;
/**
* Items
* @description Items
*/
2024-03-05 11:33:01 +00:00
items: components["schemas"]["WorkflowRecordListItemDTO"][];
2024-01-22 22:28:24 +00:00
};
/**
* Pair Tile with Image
* @description Pair an image with its tile properties.
*/
PairTileImageInvocation: {
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The tile image. */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** @description The tile properties. */
2024-03-05 11:33:01 +00:00
tile?: components["schemas"]["Tile"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* type
* @default pair_tile_image
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "pair_tile_image";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/** PairTileImageOutput */
PairTileImageOutput: {
/** @description A tile description with its corresponding image. */
2024-03-05 11:33:01 +00:00
tile_with_image: components["schemas"]["TileWithImage"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* type
* @default pair_tile_image_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "pair_tile_image_output";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* PIDI Processor
* @description Applies PIDI processing to image
*/
PidiImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Detect Resolution
* @description Pixel resolution for detection
* @default 512
*/
detect_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Image Resolution
* @description Pixel resolution for output image
* @default 512
*/
image_resolution?: number;
/**
2023-08-22 05:15:00 +00:00
* Safe
* @description Whether or not to use safe mode
* @default false
*/
safe?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Scribble
* @description Whether or not to use scribble mode
* @default false
*/
scribble?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default pidi_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "pidi_image_processor";
};
/**
2023-08-22 05:15:00 +00:00
* Prompts from File
* @description Loads prompts from a text file
*/
PromptsFromFileInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* File Path
* @description Path to prompt text file
*/
file_path?: string;
/**
2023-08-22 05:15:00 +00:00
* Pre Prompt
* @description String to prepend to each prompt
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
pre_prompt?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Post Prompt
* @description String to append to each prompt
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
post_prompt?: string | null;
/**
2023-08-22 05:15:00 +00:00
* Start Line
* @description Line in the file to start start from
* @default 1
*/
start_line?: number;
/**
2023-08-22 05:15:00 +00:00
* Max Prompts
* @description Max lines to read from file (0=all)
* @default 1
*/
max_prompts?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default prompt_from_file
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "prompt_from_file";
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* PruneResult
* @description Result of pruning the session queue
*/
PruneResult: {
/**
* Deleted
* @description Number of queue items deleted
*/
deleted: number;
};
/**
* Random Float
* @description Outputs a single random float
*/
RandomFloatInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @default false
*/
use_cache?: boolean;
/**
* Low
* @description The inclusive low value
* @default 0
*/
low?: number;
/**
* High
* @description The exclusive high value
* @default 1
*/
high?: number;
/**
* Decimals
* @description The number of decimal places to round to
* @default 2
*/
decimals?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default rand_float
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "rand_float";
};
/**
2023-08-22 05:15:00 +00:00
* Random Integer
* @description Outputs a single random integer.
*/
RandomIntInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default false
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Low
* @description The inclusive low value
* @default 0
*/
low?: number;
/**
2023-08-22 05:15:00 +00:00
* High
* @description The exclusive high value
* @default 2147483647
*/
high?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default rand_int
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "rand_int";
};
/**
2023-08-22 05:15:00 +00:00
* Random Range
* @description Creates a collection of random numbers
*/
RandomRangeInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default false
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Low
* @description The inclusive low value
* @default 0
*/
low?: number;
/**
2023-08-22 05:15:00 +00:00
* High
* @description The exclusive high value
* @default 2147483647
*/
high?: number;
/**
2023-08-22 05:15:00 +00:00
* Size
* @description The number of values to generate
* @default 1
*/
size?: number;
/**
2023-08-22 05:15:00 +00:00
* Seed
* @description The seed for the RNG (omit for random)
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default 0
*/
seed?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default random_range
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "random_range";
};
/**
2023-08-22 05:15:00 +00:00
* Integer Range
* @description Creates a range of numbers from start to stop with step
*/
RangeInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Start
* @description The start of the range
* @default 0
*/
start?: number;
/**
2023-08-22 05:15:00 +00:00
* Stop
* @description The stop of the range
* @default 10
*/
stop?: number;
/**
2023-08-22 05:15:00 +00:00
* Step
* @description The step of the range
* @default 1
*/
step?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default range
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "range";
};
/**
2023-08-22 05:15:00 +00:00
* Integer Range of Size
* @description Creates a range from start to start + (size * step) incremented by step
*/
RangeOfSizeInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Start
* @description The start of the range
* @default 0
*/
start?: number;
/**
2023-08-22 05:15:00 +00:00
* Size
* @description The number of values
* @default 1
*/
size?: number;
/**
2023-08-22 05:15:00 +00:00
* Step
* @description The step of the range
* @default 1
*/
step?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default range_of_size
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "range_of_size";
};
2024-01-22 22:28:24 +00:00
/**
* RemoteModelFile
* @description Information about a downloadable file that forms part of a model.
*/
RemoteModelFile: {
/**
* Url
* Format: uri
* @description The url to download this model file
*/
url: string;
/**
* Path
* Format: path
* @description The path to the file, relative to the model root
*/
path: string;
/**
* Size
* @description The size of this file, in bytes
*/
size: number;
/**
* Sha256
* @description SHA256 hash of this model (not always available)
*/
sha256?: string | null;
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** RemoveImagesFromBoardResult */
RemoveImagesFromBoardResult: {
/**
2023-08-22 05:15:00 +00:00
* Removed Image Names
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description The image names that were removed from their board
*/
2023-08-22 05:15:00 +00:00
removed_image_names: string[];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Resize Latents
* @description Resizes latents to explicit width/height (in pixels). Provided dimensions are floor-divided by 8.
*/
ResizeLatentsInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents?: components["schemas"]["LatentsField"];
/**
2023-08-22 05:15:00 +00:00
* Width
* @description Width of output (px)
*/
width?: number;
/**
2023-08-22 05:15:00 +00:00
* Height
* @description Width of output (px)
*/
height?: number;
/**
2023-08-22 05:15:00 +00:00
* Mode
* @description Interpolation mode
* @default bilinear
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
mode?: "nearest" | "linear" | "bilinear" | "bicubic" | "trilinear" | "area" | "nearest-exact";
2023-07-13 05:22:18 +00:00
/**
2023-08-22 05:15:00 +00:00
* Antialias
* @description Whether or not to apply antialiasing (bilinear or bicubic only)
* @default false
*/
antialias?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default lresize
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lresize";
};
/**
2023-08-22 05:15:00 +00:00
* ResourceOrigin
* @description The origin of a resource (eg image).
2023-08-22 05:15:00 +00:00
*
* - INTERNAL: The resource was created by the application.
* - EXTERNAL: The resource was not created by the application.
2023-08-22 05:15:00 +00:00
* This may be a user-initiated upload, or an internal application upload (eg Canvas init image).
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
ResourceOrigin: "internal" | "external";
/**
* Round Float
* @description Rounds a float to a specified number of decimal places.
*/
RoundInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Value
* @description The float value
* @default 0
*/
value?: number;
/**
* Decimals
* @description The number of decimal places
* @default 0
*/
decimals?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default round_float
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "round_float";
};
/**
* SDXL Prompt
* @description Parse prompt using compel package to conditioning.
*/
SDXLCompelPromptInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Prompt
* @description Prompt to be parsed by Compel to create a conditioning tensor
* @default
*/
prompt?: string;
/**
2023-08-22 05:15:00 +00:00
* Style
* @description Prompt to be parsed by Compel to create a conditioning tensor
* @default
*/
style?: string;
/**
2023-08-22 05:15:00 +00:00
* Original Width
* @default 1024
*/
original_width?: number;
/**
2023-08-22 05:15:00 +00:00
* Original Height
* @default 1024
*/
original_height?: number;
/**
2023-08-22 05:15:00 +00:00
* Crop Top
* @default 0
*/
crop_top?: number;
/**
2023-08-22 05:15:00 +00:00
* Crop Left
* @default 0
*/
crop_left?: number;
/**
2023-08-22 05:15:00 +00:00
* Target Width
* @default 1024
*/
target_width?: number;
/**
2023-08-22 05:15:00 +00:00
* Target Height
* @default 1024
*/
target_height?: number;
/**
feat: polymorphic fields Initial support for polymorphic field types. Polymorphic types are a single of or list of a specific type. For example, `Union[str, list[str]]`. Polymorphics do not yet have support for direct input in the UI (will come in the future). They will be forcibly set as Connection-only fields, in which case users will not be able to provide direct input to the field. If a polymorphic should present as a singleton type - which would allow direct input - the node must provide an explicit type hint. For example, `DenoiseLatents`' `CFG Scale` is polymorphic, but in the node editor, we want to present this as a number input. In the node definition, the field is given `ui_type=UIType.Float`, which tells the UI to treat this as a `float` field. The connection validation logic will prevent connecting a collection to `CFG Scale` in this situation, because it is typed as `float`. The workaround is to disable validation from the settings to make this specific connection. A future improvement will resolve this. This also introduces better support for collection field types. Like polymorphics, collection types are parsed automatically by the client and do not need any specific type hints. Also like polymorphics, there is no support yet for direct input of collection types in the UI. - Disabling validation in workflow editor now displays the visual hints for valid connections, but lets you connect to anything. - Added `ui_order: int` to `InputField` and `OutputField`. The UI will use this, if present, to order fields in a node UI. See usage in `DenoiseLatents` for an example. - Updated the field colors - duplicate colors have just been lightened a bit. It's not perfect but it was a quick fix. - Field handles for collections are the same color as their single counterparts, but have a dark dot in the center of them. - Field handles for polymorphics are a rounded square with dot in the middle. - Removed all fields that just render `null` from `InputFieldRenderer`, replaced with a single fallback - Removed logic in `zValidatedWorkflow`, which checked for existence of node templates for each node in a workflow. This logic introduced a circular dependency, due to importing the global redux `store` in order to get the node templates within a zod schema. It's actually fine to just leave this out entirely; The case of a missing node template is handled by the UI. Fixing it otherwise would introduce a substantial headache. - Fixed the `ControlNetInvocation.control_model` field default, which was a string when it shouldn't have one.
2023-09-01 09:40:27 +00:00
* CLIP 1
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip?: components["schemas"]["ClipField"];
/**
feat: polymorphic fields Initial support for polymorphic field types. Polymorphic types are a single of or list of a specific type. For example, `Union[str, list[str]]`. Polymorphics do not yet have support for direct input in the UI (will come in the future). They will be forcibly set as Connection-only fields, in which case users will not be able to provide direct input to the field. If a polymorphic should present as a singleton type - which would allow direct input - the node must provide an explicit type hint. For example, `DenoiseLatents`' `CFG Scale` is polymorphic, but in the node editor, we want to present this as a number input. In the node definition, the field is given `ui_type=UIType.Float`, which tells the UI to treat this as a `float` field. The connection validation logic will prevent connecting a collection to `CFG Scale` in this situation, because it is typed as `float`. The workaround is to disable validation from the settings to make this specific connection. A future improvement will resolve this. This also introduces better support for collection field types. Like polymorphics, collection types are parsed automatically by the client and do not need any specific type hints. Also like polymorphics, there is no support yet for direct input of collection types in the UI. - Disabling validation in workflow editor now displays the visual hints for valid connections, but lets you connect to anything. - Added `ui_order: int` to `InputField` and `OutputField`. The UI will use this, if present, to order fields in a node UI. See usage in `DenoiseLatents` for an example. - Updated the field colors - duplicate colors have just been lightened a bit. It's not perfect but it was a quick fix. - Field handles for collections are the same color as their single counterparts, but have a dark dot in the center of them. - Field handles for polymorphics are a rounded square with dot in the middle. - Removed all fields that just render `null` from `InputFieldRenderer`, replaced with a single fallback - Removed logic in `zValidatedWorkflow`, which checked for existence of node templates for each node in a workflow. This logic introduced a circular dependency, due to importing the global redux `store` in order to get the node templates within a zod schema. It's actually fine to just leave this out entirely; The case of a missing node template is handled by the UI. Fixing it otherwise would introduce a substantial headache. - Fixed the `ControlNetInvocation.control_model` field default, which was a string when it shouldn't have one.
2023-09-01 09:40:27 +00:00
* CLIP 2
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip2?: components["schemas"]["ClipField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default sdxl_compel_prompt
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_compel_prompt";
};
/**
2023-08-22 05:15:00 +00:00
* SDXL LoRA
* @description Apply selected lora to unet and text_encoder.
*/
SDXLLoraLoaderInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* LoRA
* @description LoRA model to load
*/
2024-03-06 08:37:52 +00:00
lora: components["schemas"]["ModelField"];
/**
2023-08-22 05:15:00 +00:00
* Weight
* @description The weight at which the LoRA is applied to each model
* @default 0.75
*/
weight?: number;
/**
feat: polymorphic fields Initial support for polymorphic field types. Polymorphic types are a single of or list of a specific type. For example, `Union[str, list[str]]`. Polymorphics do not yet have support for direct input in the UI (will come in the future). They will be forcibly set as Connection-only fields, in which case users will not be able to provide direct input to the field. If a polymorphic should present as a singleton type - which would allow direct input - the node must provide an explicit type hint. For example, `DenoiseLatents`' `CFG Scale` is polymorphic, but in the node editor, we want to present this as a number input. In the node definition, the field is given `ui_type=UIType.Float`, which tells the UI to treat this as a `float` field. The connection validation logic will prevent connecting a collection to `CFG Scale` in this situation, because it is typed as `float`. The workaround is to disable validation from the settings to make this specific connection. A future improvement will resolve this. This also introduces better support for collection field types. Like polymorphics, collection types are parsed automatically by the client and do not need any specific type hints. Also like polymorphics, there is no support yet for direct input of collection types in the UI. - Disabling validation in workflow editor now displays the visual hints for valid connections, but lets you connect to anything. - Added `ui_order: int` to `InputField` and `OutputField`. The UI will use this, if present, to order fields in a node UI. See usage in `DenoiseLatents` for an example. - Updated the field colors - duplicate colors have just been lightened a bit. It's not perfect but it was a quick fix. - Field handles for collections are the same color as their single counterparts, but have a dark dot in the center of them. - Field handles for polymorphics are a rounded square with dot in the middle. - Removed all fields that just render `null` from `InputFieldRenderer`, replaced with a single fallback - Removed logic in `zValidatedWorkflow`, which checked for existence of node templates for each node in a workflow. This logic introduced a circular dependency, due to importing the global redux `store` in order to get the node templates within a zod schema. It's actually fine to just leave this out entirely; The case of a missing node template is handled by the UI. Fixing it otherwise would introduce a substantial headache. - Fixed the `ControlNetInvocation.control_model` field default, which was a string when it shouldn't have one.
2023-09-01 09:40:27 +00:00
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet?: components["schemas"]["UNetField"] | null;
/**
2023-08-22 05:15:00 +00:00
* CLIP 1
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip?: components["schemas"]["ClipField"] | null;
/**
2023-08-22 05:15:00 +00:00
* CLIP 2
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip2?: components["schemas"]["ClipField"] | null;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default sdxl_lora_loader
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_lora_loader";
};
/**
2023-08-22 05:15:00 +00:00
* SDXLLoraLoaderOutput
* @description SDXL LoRA Loader Output
*/
SDXLLoraLoaderOutput: {
/**
2023-08-22 05:15:00 +00:00
* UNet
* @description UNet (scheduler, LoRAs)
2024-02-17 09:02:51 +00:00
* @default null
*/
2024-03-05 11:33:01 +00:00
unet: components["schemas"]["UNetField"] | null;
/**
2023-08-22 05:15:00 +00:00
* CLIP 1
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
2024-02-17 09:02:51 +00:00
* @default null
*/
2024-03-05 11:33:01 +00:00
clip: components["schemas"]["ClipField"] | null;
/**
2023-08-22 05:15:00 +00:00
* CLIP 2
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
2024-02-17 09:02:51 +00:00
* @default null
*/
2024-03-05 11:33:01 +00:00
clip2: components["schemas"]["ClipField"] | null;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default sdxl_lora_loader_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_lora_loader_output";
};
/**
2023-08-22 05:15:00 +00:00
* SDXL Main Model
* @description Loads an sdxl base model, outputting its submodels.
*/
SDXLModelLoaderInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description SDXL Main model (UNet, VAE, CLIP1, CLIP2) to load */
2024-03-06 08:37:52 +00:00
model: components["schemas"]["ModelField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default sdxl_model_loader
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_model_loader";
};
/**
2023-08-22 05:15:00 +00:00
* SDXLModelLoaderOutput
* @description SDXL base model loader output
*/
SDXLModelLoaderOutput: {
/**
2023-08-22 05:15:00 +00:00
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet: components["schemas"]["UNetField"];
/**
2023-08-22 05:15:00 +00:00
* CLIP 1
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip: components["schemas"]["ClipField"];
/**
2023-08-22 05:15:00 +00:00
* CLIP 2
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip2: components["schemas"]["ClipField"];
/**
2023-08-22 05:15:00 +00:00
* VAE
* @description VAE
*/
2024-03-05 11:33:01 +00:00
vae: components["schemas"]["VaeField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default sdxl_model_loader_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_model_loader_output";
};
/**
* SDXL Refiner Prompt
* @description Parse prompt using compel package to conditioning.
*/
SDXLRefinerCompelPromptInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Style
* @description Prompt to be parsed by Compel to create a conditioning tensor
* @default
*/
style?: string;
/**
2023-08-22 05:15:00 +00:00
* Original Width
* @default 1024
*/
original_width?: number;
/**
2023-08-22 05:15:00 +00:00
* Original Height
* @default 1024
*/
original_height?: number;
/**
2023-08-22 05:15:00 +00:00
* Crop Top
* @default 0
*/
crop_top?: number;
/**
2023-08-22 05:15:00 +00:00
* Crop Left
* @default 0
*/
crop_left?: number;
/**
2023-08-22 05:15:00 +00:00
* Aesthetic Score
* @description The aesthetic score to apply to the conditioning tensor
* @default 6
*/
aesthetic_score?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count */
2024-03-05 11:33:01 +00:00
clip2?: components["schemas"]["ClipField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default sdxl_refiner_compel_prompt
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_refiner_compel_prompt";
};
/**
2023-08-22 05:15:00 +00:00
* SDXL Refiner Model
* @description Loads an sdxl refiner model, outputting its submodels.
*/
SDXLRefinerModelLoaderInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description SDXL Refiner Main Modde (UNet, VAE, CLIP2) to load */
2024-03-06 08:37:52 +00:00
model: components["schemas"]["ModelField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default sdxl_refiner_model_loader
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_refiner_model_loader";
};
/**
2023-08-22 05:15:00 +00:00
* SDXLRefinerModelLoaderOutput
* @description SDXL refiner model loader output
*/
SDXLRefinerModelLoaderOutput: {
/**
2023-08-22 05:15:00 +00:00
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet: components["schemas"]["UNetField"];
/**
2023-08-22 05:15:00 +00:00
* CLIP 2
* @description CLIP (tokenizer, text encoder, LoRAs) and skipped layer count
*/
2024-03-05 11:33:01 +00:00
clip2: components["schemas"]["ClipField"];
/**
2023-08-22 05:15:00 +00:00
* VAE
* @description VAE
*/
2024-03-05 11:33:01 +00:00
vae: components["schemas"]["VaeField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default sdxl_refiner_model_loader_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sdxl_refiner_model_loader_output";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* SQLiteDirection
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
SQLiteDirection: "ASC" | "DESC";
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Save Image
* @description Saves an image. Unlike an image primitive, this invocation stores a copy of the image.
*/
SaveImageInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default false
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
* @default save_image
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "save_image";
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Scale Latents
* @description Scales latents by a given factor.
*/
ScaleLatentsInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Latents tensor */
2024-03-05 11:33:01 +00:00
latents?: components["schemas"]["LatentsField"];
/**
2023-08-22 05:15:00 +00:00
* Scale Factor
* @description The factor by which to scale
*/
scale_factor?: number;
/**
2023-08-22 05:15:00 +00:00
* Mode
* @description Interpolation mode
* @default bilinear
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
mode?: "nearest" | "linear" | "bilinear" | "bicubic" | "trilinear" | "area" | "nearest-exact";
2023-07-13 05:22:18 +00:00
/**
2023-08-22 05:15:00 +00:00
* Antialias
* @description Whether or not to apply antialiasing (bilinear or bicubic only)
* @default false
*/
antialias?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default lscale
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "lscale";
};
2023-09-02 01:36:07 +00:00
/**
* Scheduler
* @description Selects a scheduler.
*/
SchedulerInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
2023-09-02 01:36:07 +00:00
/**
* Scheduler
* @description Scheduler to use during inference
* @default euler
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
scheduler?: "ddim" | "ddpm" | "deis" | "lms" | "lms_k" | "pndm" | "heun" | "heun_k" | "euler" | "euler_k" | "euler_a" | "kdpm_2" | "kdpm_2_a" | "dpmpp_2s" | "dpmpp_2s_k" | "dpmpp_2m" | "dpmpp_2m_k" | "dpmpp_2m_sde" | "dpmpp_2m_sde_k" | "dpmpp_sde" | "dpmpp_sde_k" | "unipc" | "lcm";
2023-09-02 01:36:07 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-09-02 01:36:07 +00:00
* @default scheduler
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-09-02 01:36:07 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "scheduler";
2023-09-02 01:36:07 +00:00
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** SchedulerOutput */
2023-09-02 01:36:07 +00:00
SchedulerOutput: {
/**
* Scheduler
* @description Scheduler to use during inference
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
scheduler: "ddim" | "ddpm" | "deis" | "lms" | "lms_k" | "pndm" | "heun" | "heun_k" | "euler" | "euler_k" | "euler_a" | "kdpm_2" | "kdpm_2_a" | "dpmpp_2s" | "dpmpp_2s_k" | "dpmpp_2m" | "dpmpp_2m_k" | "dpmpp_2m_sde" | "dpmpp_2m_sde_k" | "dpmpp_sde" | "dpmpp_sde_k" | "unipc" | "lcm";
2023-09-02 01:36:07 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-09-02 01:36:07 +00:00
* @default scheduler_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-09-02 01:36:07 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "scheduler_output";
2023-09-02 01:36:07 +00:00
};
2024-02-15 11:15:21 +00:00
/**
* SchedulerPredictionType
* @description Scheduler prediction type.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
SchedulerPredictionType: "epsilon" | "v_prediction" | "sample";
2023-08-29 00:20:55 +00:00
/**
* Seamless
* @description Applies the seamless transformation to the Model UNet and VAE.
*/
SeamlessModeInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-08-29 00:20:55 +00:00
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-08-29 00:20:55 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
2023-08-29 00:20:55 +00:00
/**
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet?: components["schemas"]["UNetField"] | null;
2023-08-29 00:20:55 +00:00
/**
* VAE
* @description VAE model to load
*/
2024-03-05 11:33:01 +00:00
vae?: components["schemas"]["VaeField"] | null;
2023-08-29 00:20:55 +00:00
/**
* Seamless Y
* @description Specify whether Y axis is seamless
* @default true
*/
seamless_y?: boolean;
/**
* Seamless X
* @description Specify whether X axis is seamless
* @default true
*/
seamless_x?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default seamless
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "seamless";
2023-08-29 00:20:55 +00:00
};
/**
* SeamlessModeOutput
* @description Modified Seamless Model output
*/
SeamlessModeOutput: {
/**
* UNet
* @description UNet (scheduler, LoRAs)
2024-02-17 09:02:51 +00:00
* @default null
2023-08-29 00:20:55 +00:00
*/
2024-03-05 11:33:01 +00:00
unet: components["schemas"]["UNetField"] | null;
2023-08-29 00:20:55 +00:00
/**
* VAE
* @description VAE
2024-02-17 09:02:51 +00:00
* @default null
2023-08-29 00:20:55 +00:00
*/
2024-03-05 11:33:01 +00:00
vae: components["schemas"]["VaeField"] | null;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default seamless_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "seamless_output";
2023-08-29 00:20:55 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Segment Anything Processor
* @description Applies segment anything processing to image
*/
SegmentAnythingProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default segment_anything_processor
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "segment_anything_processor";
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** SessionProcessorStatus */
SessionProcessorStatus: {
/**
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
* Is Started
* @description Whether the session processor is started
*/
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
is_started: boolean;
/**
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
* Is Processing
* @description Whether a session is being processed
*/
is_processing: boolean;
};
/**
* SessionQueueAndProcessorStatus
* @description The overall status of session queue and processor
*/
SessionQueueAndProcessorStatus: {
2024-03-05 11:33:01 +00:00
queue: components["schemas"]["SessionQueueStatus"];
processor: components["schemas"]["SessionProcessorStatus"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** SessionQueueItem */
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
SessionQueueItem: {
/**
* Item Id
* @description The identifier of the session queue item
*/
item_id: number;
/**
* Status
* @description The status of this queue item
* @default pending
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
status: "pending" | "in_progress" | "completed" | "failed" | "canceled";
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Priority
* @description The priority of this queue item
* @default 0
*/
priority: number;
/**
* Batch Id
* @description The ID of the batch associated with this queue item
*/
batch_id: string;
/**
* Session Id
* @description The ID of the session associated with this queue item. The session doesn't exist in graph_executions until the queue item is executed.
*/
session_id: string;
/**
* Error
* @description The error message if this queue item errored
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
error?: string | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Created At
* @description When this queue item was created
*/
created_at: string;
/**
* Updated At
* @description When this queue item was updated
*/
updated_at: string;
/**
* Started At
* @description When this queue item was started
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
started_at?: string | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Completed At
* @description When this queue item was completed
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
completed_at?: string | null;
/**
* Queue Id
* @description The id of the queue with which this item is associated
*/
queue_id: string;
/**
* Field Values
* @description The field values that were used for this queue item
*/
2024-03-05 11:33:01 +00:00
field_values?: components["schemas"]["NodeFieldValue"][] | null;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The fully-populated session to be executed */
2024-03-05 11:33:01 +00:00
session: components["schemas"]["GraphExecutionState"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** @description The workflow associated with this queue item */
2024-03-05 11:33:01 +00:00
workflow?: components["schemas"]["WorkflowWithoutID"] | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** SessionQueueItemDTO */
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
SessionQueueItemDTO: {
/**
* Item Id
* @description The identifier of the session queue item
*/
item_id: number;
/**
* Status
* @description The status of this queue item
* @default pending
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
status: "pending" | "in_progress" | "completed" | "failed" | "canceled";
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Priority
* @description The priority of this queue item
* @default 0
*/
priority: number;
/**
* Batch Id
* @description The ID of the batch associated with this queue item
*/
batch_id: string;
/**
* Session Id
* @description The ID of the session associated with this queue item. The session doesn't exist in graph_executions until the queue item is executed.
*/
session_id: string;
/**
* Error
* @description The error message if this queue item errored
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
error?: string | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Created At
* @description When this queue item was created
*/
created_at: string;
/**
* Updated At
* @description When this queue item was updated
*/
updated_at: string;
/**
* Started At
* @description When this queue item was started
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
started_at?: string | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Completed At
* @description When this queue item was completed
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
completed_at?: string | null;
/**
* Queue Id
* @description The id of the queue with which this item is associated
*/
queue_id: string;
/**
* Field Values
* @description The field values that were used for this queue item
*/
2024-03-05 11:33:01 +00:00
field_values?: components["schemas"]["NodeFieldValue"][] | null;
2024-02-09 20:14:15 +00:00
};
/** SessionQueueStatus */
SessionQueueStatus: {
/**
* Queue Id
* @description The ID of the queue
*/
queue_id: string;
/**
* Item Id
* @description The current queue item id
*/
item_id: number | null;
/**
* Batch Id
* @description The current queue item's batch id
*/
batch_id: string | null;
/**
* Session Id
* @description The current queue item's session id
*/
session_id: string | null;
/**
* Pending
* @description Number of queue items with status 'pending'
*/
pending: number;
/**
* In Progress
* @description Number of queue items with status 'in_progress'
*/
in_progress: number;
/**
* Completed
* @description Number of queue items with status 'complete'
*/
completed: number;
/**
* Failed
* @description Number of queue items with status 'error'
*/
failed: number;
/**
* Canceled
* @description Number of queue items with status 'canceled'
*/
canceled: number;
/**
* Total
* @description Total number of queue items
*/
total: number;
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Show Image
* @description Displays a provided image using the OS image viewer, and passes it forward in the pipeline.
*/
ShowImageInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to show */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default show_image
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "show_image";
};
/**
2023-08-22 05:15:00 +00:00
* Step Param Easing
* @description Experimental per-step parameter easing for denoising steps
*/
StepParamEasingInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
2023-08-16 01:59:19 +00:00
/**
2023-08-22 05:15:00 +00:00
* Easing
* @description The easing function to use
* @default Linear
2023-08-16 01:59:19 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
easing?: "Linear" | "QuadIn" | "QuadOut" | "QuadInOut" | "CubicIn" | "CubicOut" | "CubicInOut" | "QuarticIn" | "QuarticOut" | "QuarticInOut" | "QuinticIn" | "QuinticOut" | "QuinticInOut" | "SineIn" | "SineOut" | "SineInOut" | "CircularIn" | "CircularOut" | "CircularInOut" | "ExponentialIn" | "ExponentialOut" | "ExponentialInOut" | "ElasticIn" | "ElasticOut" | "ElasticInOut" | "BackIn" | "BackOut" | "BackInOut" | "BounceIn" | "BounceOut" | "BounceInOut";
2023-08-16 01:59:19 +00:00
/**
2023-08-22 05:15:00 +00:00
* Num Steps
* @description number of denoising steps
* @default 20
*/
num_steps?: number;
/**
2023-08-22 05:15:00 +00:00
* Start Value
* @description easing starting value
* @default 0
*/
start_value?: number;
/**
2023-08-22 05:15:00 +00:00
* End Value
* @description easing ending value
* @default 1
*/
end_value?: number;
/**
2023-08-22 05:15:00 +00:00
* Start Step Percent
* @description fraction of steps at which to start easing
* @default 0
*/
start_step_percent?: number;
/**
2023-08-22 05:15:00 +00:00
* End Step Percent
* @description fraction of steps after which to end easing
* @default 1
*/
end_step_percent?: number;
/**
2023-08-22 05:15:00 +00:00
* Pre Start Value
* @description value before easing start
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
pre_start_value?: number | null;
/**
2023-08-22 05:15:00 +00:00
* Post End Value
* @description value after easing end
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
post_end_value?: number | null;
/**
2023-08-22 05:15:00 +00:00
* Mirror
* @description include mirror of easing function
* @default false
*/
mirror?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Show Easing Plot
* @description show easing plot
* @default false
*/
show_easing_plot?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default step_param_easing
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "step_param_easing";
};
/**
* String2Output
* @description Base class for invocations that output two strings
*/
String2Output: {
/**
* String 1
* @description string 1
*/
string_1: string;
/**
* String 2
* @description string 2
*/
string_2: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string_2_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_2_output";
};
/**
* String Collection Primitive
* @description A collection of string primitive values
*/
StringCollectionInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* Collection
* @description The collection of string values
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @default []
*/
collection?: string[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default string_collection
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_collection";
};
/**
2023-08-22 05:15:00 +00:00
* StringCollectionOutput
* @description Base class for nodes that output a collection of strings
*/
StringCollectionOutput: {
/**
* Collection
* @description The output strings
*/
collection: string[];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default string_collection_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_collection_output";
};
/**
2023-08-22 05:15:00 +00:00
* String Primitive
* @description A string primitive value
*/
StringInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* Value
* @description The string value
* @default
*/
value?: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string";
};
/**
* String Join
* @description Joins string left to string right
*/
StringJoinInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* String Left
* @description String Left
* @default
*/
string_left?: string;
/**
* String Right
* @description String Right
* @default
*/
string_right?: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string_join
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_join";
};
/**
* String Join Three
* @description Joins string left to string middle to string right
*/
StringJoinThreeInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* String Left
* @description String Left
* @default
*/
string_left?: string;
/**
* String Middle
* @description String Middle
* @default
*/
string_middle?: string;
/**
* String Right
* @description String Right
* @default
*/
string_right?: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string_join_three
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_join_three";
};
/**
2023-08-22 05:15:00 +00:00
* StringOutput
* @description Base class for nodes that output a single string
*/
StringOutput: {
/**
* Value
* @description The output string
*/
value: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default string_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_output";
};
/**
* StringPosNegOutput
* @description Base class for invocations that output a positive and negative string
*/
StringPosNegOutput: {
/**
* Positive String
* @description Positive string
*/
positive_string: string;
/**
* Negative String
* @description Negative string
*/
negative_string: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string_pos_neg_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_pos_neg_output";
};
/**
* String Replace
* @description Replaces the search string with the replace string
*/
StringReplaceInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* String
* @description String to work on
* @default
*/
string?: string;
/**
* Search String
* @description String to search for
* @default
*/
search_string?: string;
/**
* Replace String
* @description String to replace the search
* @default
*/
replace_string?: string;
/**
* Use Regex
* @description Use search string as a regex expression (non regex is case insensitive)
* @default false
*/
use_regex?: boolean;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string_replace
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_replace";
};
/**
* String Split
* @description Splits string into two strings, based on the first occurance of the delimiter. The delimiter will be removed from the string
*/
StringSplitInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* String
* @description String to split
* @default
*/
string?: string;
/**
* Delimiter
* @description Delimiter to spilt with. blank will split on the first whitespace
* @default
*/
delimiter?: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string_split
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_split";
};
/**
* String Split Negative
* @description Splits string into two strings, inside [] goes into negative string everthing else goes into positive string. Each [ and ] character is replaced with a space
*/
StringSplitNegInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* String
* @description String to split
* @default
*/
string?: string;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default string_split_neg
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "string_split_neg";
};
/**
2023-08-22 05:15:00 +00:00
* SubModelType
2024-02-15 11:15:21 +00:00
* @description Submodel type.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
SubModelType: "unet" | "text_encoder" | "text_encoder_2" | "tokenizer" | "tokenizer_2" | "vae" | "vae_decoder" | "vae_encoder" | "scheduler" | "safety_checker";
2023-08-16 01:59:19 +00:00
/**
2023-08-22 05:15:00 +00:00
* Subtract Integers
* @description Subtracts two numbers
*/
SubtractInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
2023-08-22 05:15:00 +00:00
* A
* @description The first number
* @default 0
*/
a?: number;
/**
2023-08-22 05:15:00 +00:00
* B
* @description The second number
* @default 0
*/
b?: number;
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default sub
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "sub";
};
/**
* T2IAdapterConfig
* @description Model config for T2I.
*/
T2IAdapterConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
*/
path: string;
/**
* Name
* @description Name of the model.
*/
name: string;
/** @description The base model. */
base: components["schemas"]["BaseModelType"];
/**
* Description
* @description Model description
*/
description?: string | null;
/**
* Source
* @description The original source of the model (path, URL or repo_id).
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
*/
source_api_response?: string | null;
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
/**
* Type
* @default t2i_adapter
* @constant
*/
type: "t2i_adapter";
/**
* Format
* @constant
*/
format: "diffusers";
};
/** T2IAdapterField */
T2IAdapterField: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The T2I-Adapter image prompt. */
2024-03-05 11:33:01 +00:00
image: components["schemas"]["ImageField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The T2I-Adapter model to use. */
2024-03-06 08:37:52 +00:00
t2i_adapter_model: components["schemas"]["ModelField"];
/**
* Weight
* @description The weight given to the T2I-Adapter
* @default 1
*/
weight?: number | number[];
/**
* Begin Step Percent
* @description When the T2I-Adapter is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
* End Step Percent
* @description When the T2I-Adapter is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
/**
* Resize Mode
* @description The resize mode to use
* @default just_resize
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
resize_mode?: "just_resize" | "crop_resize" | "fill_resize" | "just_resize_simple";
};
/**
* T2I-Adapter
* @description Collects T2I-Adapter info to pass to other nodes.
*/
T2IAdapterInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The IP-Adapter image prompt. */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* T2I-Adapter Model
* @description The T2I-Adapter model.
*/
2024-03-06 08:37:52 +00:00
t2i_adapter_model: components["schemas"]["ModelField"];
/**
* Weight
* @description The weight given to the T2I-Adapter
* @default 1
*/
weight?: number | number[];
/**
* Begin Step Percent
* @description When the T2I-Adapter is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
* End Step Percent
* @description When the T2I-Adapter is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
/**
* Resize Mode
* @description The resize mode applied to the T2I-Adapter input image so that it matches the target output size.
* @default just_resize
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
resize_mode?: "just_resize" | "crop_resize" | "fill_resize" | "just_resize_simple";
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default t2i_adapter
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "t2i_adapter";
};
2024-03-06 08:37:52 +00:00
/** T2IAdapterMetadataField */
T2IAdapterMetadataField: {
/** @description The T2I-Adapter image prompt. */
image: components["schemas"]["ImageField"];
/** @description The T2I-Adapter model to use. */
t2i_adapter_model: components["schemas"]["ModelMetadataField"];
/**
2024-03-06 08:37:52 +00:00
* Weight
* @description The weight given to the T2I-Adapter
* @default 1
*/
2024-03-06 08:37:52 +00:00
weight?: number | number[];
/**
* Begin Step Percent
* @description When the T2I-Adapter is first applied (% of total steps)
* @default 0
*/
begin_step_percent?: number;
/**
* End Step Percent
* @description When the T2I-Adapter is last applied (% of total steps)
* @default 1
*/
end_step_percent?: number;
/**
* Resize Mode
* @description The resize mode to use
* @default just_resize
* @enum {string}
*/
resize_mode?: "just_resize" | "crop_resize" | "fill_resize" | "just_resize_simple";
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** T2IAdapterOutput */
T2IAdapterOutput: {
/**
* T2I Adapter
* @description T2I-Adapter(s) to apply
*/
2024-03-05 11:33:01 +00:00
t2i_adapter: components["schemas"]["T2IAdapterField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
* @default t2i_adapter_output
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "t2i_adapter_output";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** TBLR */
TBLR: {
/** Top */
top: number;
/** Bottom */
bottom: number;
/** Left */
left: number;
/** Right */
right: number;
};
2023-11-16 00:21:28 +00:00
/**
* TextualInversionFileConfig
2023-11-16 00:21:28 +00:00
* @description Model config for textual inversion embeddings.
*/
TextualInversionFileConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
/**
* Description
* @description Model description
*/
description?: string | null;
/**
* Source
* @description The original source of the model (path, URL or repo_id).
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
*/
source_api_response?: string | null;
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2023-11-16 00:21:28 +00:00
/**
* Type
* @default embedding
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "embedding";
2023-11-16 00:21:28 +00:00
/**
* Format
* @default embedding_file
* @constant
2023-11-16 00:21:28 +00:00
*/
format: "embedding_file";
};
/**
* TextualInversionFolderConfig
* @description Model config for textual inversion embeddings.
*/
TextualInversionFolderConfig: {
2023-11-16 00:21:28 +00:00
/**
* Key
* @description A unique key for this model.
2023-11-16 00:21:28 +00:00
*/
key: string;
2023-11-16 00:21:28 +00:00
/**
* Hash
* @description The hash of the model file(s).
2023-11-16 00:21:28 +00:00
*/
hash: string;
2023-11-16 00:21:28 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2023-11-16 00:21:28 +00:00
*/
path: string;
/**
* Name
* @description Name of the model.
*/
name: string;
/** @description The base model. */
base: components["schemas"]["BaseModelType"];
2024-02-15 11:15:21 +00:00
/**
* Description
* @description Model description
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
description?: string | null;
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
*/
source_api_response?: string | null;
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
/**
* Type
* @default embedding
* @constant
*/
type: "embedding";
/**
* Format
* @default embedding_folder
* @constant
*/
format: "embedding_folder";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** Tile */
Tile: {
/** @description The coordinates of this tile relative to its parent image. */
2024-03-05 11:33:01 +00:00
coords: components["schemas"]["TBLR"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** @description The amount of overlap with adjacent tiles on each side of this tile. */
2024-03-05 11:33:01 +00:00
overlap: components["schemas"]["TBLR"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Tile Resample Processor
* @description Tile resampler processor
*/
TileResamplerProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
2023-08-22 05:15:00 +00:00
* Down Sampling Rate
* @description Down sampling rate
* @default 1
*/
down_sampling_rate?: number;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* type
* @default tile_image_processor
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "tile_image_processor";
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Tile to Properties
* @description Split a Tile into its individual properties.
*/
TileToPropertiesInvocation: {
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The tile to split into properties. */
2024-03-05 11:33:01 +00:00
tile?: components["schemas"]["Tile"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* type
* @default tile_to_properties
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "tile_to_properties";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/** TileToPropertiesOutput */
TileToPropertiesOutput: {
/**
* Coords Left
* @description Left coordinate of the tile relative to its parent image.
*/
coords_left: number;
/**
* Coords Right
* @description Right coordinate of the tile relative to its parent image.
*/
coords_right: number;
/**
* Coords Top
* @description Top coordinate of the tile relative to its parent image.
*/
coords_top: number;
/**
* Coords Bottom
* @description Bottom coordinate of the tile relative to its parent image.
*/
coords_bottom: number;
/**
* Width
* @description The width of the tile. Equal to coords_right - coords_left.
*/
width: number;
/**
* Height
* @description The height of the tile. Equal to coords_bottom - coords_top.
*/
height: number;
/**
* Overlap Top
* @description Overlap between this tile and its top neighbor.
*/
overlap_top: number;
/**
* Overlap Bottom
* @description Overlap between this tile and its bottom neighbor.
*/
overlap_bottom: number;
/**
* Overlap Left
* @description Overlap between this tile and its left neighbor.
*/
overlap_left: number;
/**
* Overlap Right
* @description Overlap between this tile and its right neighbor.
*/
overlap_right: number;
/**
* type
* @default tile_to_properties_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "tile_to_properties_output";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/** TileWithImage */
TileWithImage: {
2024-03-05 11:33:01 +00:00
tile: components["schemas"]["Tile"];
image: components["schemas"]["ImageField"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/** UNetField */
UNetField: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Info to load unet submodel */
2024-03-06 08:37:52 +00:00
unet: components["schemas"]["ModelField"];
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Info to load scheduler submodel */
2024-03-06 08:37:52 +00:00
scheduler: components["schemas"]["ModelField"];
/**
2023-08-22 05:15:00 +00:00
* Loras
* @description Loras to apply on model loading
*/
2024-03-06 08:37:52 +00:00
loras: components["schemas"]["LoRAField"][];
2023-08-29 00:20:55 +00:00
/**
* Seamless Axes
* @description Axes("x" and "y") to which apply seamless
*/
2023-09-04 05:25:58 +00:00
seamless_axes?: string[];
2024-02-17 09:02:51 +00:00
/**
* @description FreeU configuration
* @default null
*/
2024-03-05 11:33:01 +00:00
freeu_config?: components["schemas"]["FreeUConfig"] | null;
2023-11-08 06:28:37 +00:00
};
/**
* UNetOutput
2024-02-15 11:15:21 +00:00
* @description Base class for invocations that output a UNet field.
2023-11-08 06:28:37 +00:00
*/
UNetOutput: {
/**
* UNet
* @description UNet (scheduler, LoRAs)
*/
2024-03-05 11:33:01 +00:00
unet: components["schemas"]["UNetField"];
2023-11-08 06:28:37 +00:00
/**
* type
* @default unet_output
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "unet_output";
};
/**
* URLModelSource
* @description A generic URL point to a checkpoint file.
*/
URLModelSource: {
/**
* Url
* Format: uri
*/
url: string;
/** Access Token */
access_token?: string | null;
/**
* Type
2024-01-22 22:28:24 +00:00
* @default url
* @constant
*/
2024-03-05 11:33:01 +00:00
type?: "url";
};
/**
* Unsharp Mask
* @description Applies an unsharp mask filter to an image
*/
UnsharpMaskInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
is_intermediate?: boolean;
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/** @description The image to use */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
* Radius
* @description Unsharp mask radius
* @default 2
*/
radius?: number;
/**
* Strength
* @description Unsharp mask strength
* @default 50
*/
strength?: number;
/**
* type
* @default unsharp_mask
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "unsharp_mask";
};
/** Upscaler */
Upscaler: {
/**
2023-08-22 05:15:00 +00:00
* Upscaling Method
* @description Name of upscaling method
*/
upscaling_method: string;
/**
2023-08-22 05:15:00 +00:00
* Upscaling Models
* @description List of upscaling models for this method
*/
2023-08-22 05:15:00 +00:00
upscaling_models: string[];
};
2023-06-30 22:15:04 +00:00
/**
* VAECheckpointConfig
* @description Model config for standalone VAE models.
2023-06-30 22:15:04 +00:00
*/
VAECheckpointConfig: {
2023-06-30 22:15:04 +00:00
/**
2024-02-15 11:15:21 +00:00
* Key
* @description A unique key for this model.
2023-06-30 22:15:04 +00:00
*/
2024-02-15 11:15:21 +00:00
key: string;
2023-11-08 06:28:37 +00:00
/**
* Hash
* @description The hash of the model file(s).
2023-11-08 06:28:37 +00:00
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
2023-11-16 00:21:28 +00:00
/**
* Description
* @description Model description
2023-11-16 00:21:28 +00:00
*/
description?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Source
* @description The original source of the model (path, URL or repo_id).
2023-11-16 00:21:28 +00:00
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
2023-11-16 00:21:28 +00:00
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
2023-11-16 00:21:28 +00:00
*/
source_api_response?: string | null;
2023-11-16 00:21:28 +00:00
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
2023-11-16 00:21:28 +00:00
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2023-11-16 00:21:28 +00:00
/**
* Format
* @default checkpoint
* @constant
2023-11-16 00:21:28 +00:00
*/
format: "checkpoint";
2024-02-15 11:15:21 +00:00
/**
* Config Path
* @description path to the checkpoint model config file
2024-02-15 11:15:21 +00:00
*/
config_path: string;
2023-11-16 00:21:28 +00:00
/**
* Converted At
* @description When this model was last converted to diffusers
2023-11-16 00:21:28 +00:00
*/
converted_at?: number | null;
2024-02-15 11:15:21 +00:00
/**
* Type
* @default vae
* @constant
2024-02-15 11:15:21 +00:00
*/
type: "vae";
2023-11-16 00:21:28 +00:00
};
/**
* VAEDiffusersConfig
2023-11-16 00:21:28 +00:00
* @description Model config for standalone VAE models (diffusers version).
*/
VAEDiffusersConfig: {
/**
* Key
* @description A unique key for this model.
*/
key: string;
/**
* Hash
* @description The hash of the model file(s).
*/
hash: string;
2024-02-15 11:15:21 +00:00
/**
* Path
* @description Path to the model on the filesystem. Relative paths are relative to the Invoke root directory.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
path: string;
2024-02-15 11:15:21 +00:00
/**
* Name
* @description Name of the model.
2024-02-15 11:15:21 +00:00
*/
2023-11-16 00:21:28 +00:00
name: string;
/** @description The base model. */
2024-02-15 11:15:21 +00:00
base: components["schemas"]["BaseModelType"];
/**
* Description
* @description Model description
*/
description?: string | null;
/**
* Source
* @description The original source of the model (path, URL or repo_id).
*/
source: string;
/** @description The type of source */
source_type: components["schemas"]["ModelSourceType"];
/**
* Source Api Response
* @description The original API response from the source, as stringified JSON.
*/
source_api_response?: string | null;
/**
* Trigger Phrases
* @description Set of trigger phrases for this model
*/
trigger_phrases?: string[] | null;
/** @description Default settings for this model */
default_settings?: components["schemas"]["ModelDefaultSettings"] | null;
2023-11-16 00:21:28 +00:00
/**
* Type
* @default vae
* @constant
*/
type: "vae";
2023-11-16 00:21:28 +00:00
/**
* Format
* @default diffusers
* @constant
*/
format: "diffusers";
};
/**
* VAEOutput
* @description Base class for invocations that output a VAE field
*/
VAEOutput: {
2023-11-16 00:21:28 +00:00
/**
* VAE
* @description VAE
2023-11-16 00:21:28 +00:00
*/
2024-03-05 11:33:01 +00:00
vae: components["schemas"]["VaeField"];
2024-02-15 11:15:21 +00:00
/**
* type
* @default vae_output
* @constant
2024-02-15 11:15:21 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "vae_output";
2023-11-16 00:21:28 +00:00
};
/** VaeField */
VaeField: {
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description Info to load vae submodel */
2024-03-06 08:37:52 +00:00
vae: components["schemas"]["ModelField"];
2023-08-29 00:20:55 +00:00
/**
* Seamless Axes
* @description Axes("x" and "y") to which apply seamless
*/
2023-09-04 05:25:58 +00:00
seamless_axes?: string[];
};
2023-06-30 22:15:04 +00:00
/**
2023-08-22 05:15:00 +00:00
* VAE
2023-06-30 22:15:04 +00:00
* @description Loads a VAE model, outputting a VaeLoaderOutput
*/
VaeLoaderInvocation: {
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
2023-06-30 22:15:04 +00:00
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
2023-06-30 22:15:04 +00:00
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
/**
* VAE
* @description VAE model to load
*/
2024-03-06 08:37:52 +00:00
vae_model: components["schemas"]["ModelField"];
2023-06-30 22:15:04 +00:00
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default vae_loader
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
2023-06-30 22:15:04 +00:00
*/
2024-03-05 11:33:01 +00:00
type: "vae_loader";
2023-06-30 22:15:04 +00:00
};
/** ValidationError */
ValidationError: {
/** Location */
loc: (string | number)[];
/** Message */
msg: string;
/** Error Type */
type: string;
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** Workflow */
Workflow: {
/**
* Name
* @description The name of the workflow.
*/
name: string;
/**
* Author
* @description The author of the workflow.
*/
author: string;
/**
* Description
* @description The description of the workflow.
*/
description: string;
/**
* Version
* @description The version of the workflow.
*/
version: string;
/**
* Contact
* @description The contact of the workflow.
*/
contact: string;
/**
* Tags
* @description The tags of the workflow.
*/
tags: string;
/**
* Notes
* @description The notes of the workflow.
*/
notes: string;
/**
* Exposedfields
* @description The exposed fields of the workflow.
*/
2024-03-05 11:33:01 +00:00
exposedFields: components["schemas"]["ExposedField"][];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** @description The meta of the workflow. */
2024-03-05 11:33:01 +00:00
meta: components["schemas"]["WorkflowMeta"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Nodes
* @description The nodes of the workflow.
*/
nodes: {
2024-03-05 11:33:01 +00:00
[key: string]: components["schemas"]["JsonValue"];
}[];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Edges
* @description The edges of the workflow.
*/
edges: {
2024-03-05 11:33:01 +00:00
[key: string]: components["schemas"]["JsonValue"];
}[];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Id
* @description The id of the workflow.
*/
2023-12-06 12:33:54 +00:00
id: string;
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/**
* WorkflowCategory
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
WorkflowCategory: "user" | "default" | "project";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** WorkflowMeta */
WorkflowMeta: {
/**
* Version
* @description The version of the workflow schema.
*/
version: string;
2023-12-06 12:33:54 +00:00
/**
* @description The category of the workflow (user or default).
* @default user
*/
2024-03-05 11:33:01 +00:00
category?: components["schemas"]["WorkflowCategory"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/** WorkflowRecordDTO */
WorkflowRecordDTO: {
/**
* Workflow Id
* @description The id of the workflow.
*/
workflow_id: string;
/**
* Name
* @description The name of the workflow.
*/
name: string;
/**
* Created At
* @description The created timestamp of the workflow.
*/
created_at: string;
/**
* Updated At
* @description The updated timestamp of the workflow.
*/
updated_at: string;
/**
* Opened At
* @description The opened timestamp of the workflow.
*/
opened_at: string;
/** @description The workflow. */
2024-03-05 11:33:01 +00:00
workflow: components["schemas"]["Workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/** WorkflowRecordListItemDTO */
WorkflowRecordListItemDTO: {
/**
* Workflow Id
* @description The id of the workflow.
*/
workflow_id: string;
/**
* Name
* @description The name of the workflow.
*/
name: string;
/**
* Created At
* @description The created timestamp of the workflow.
*/
created_at: string;
/**
* Updated At
* @description The updated timestamp of the workflow.
*/
updated_at: string;
/**
* Opened At
* @description The opened timestamp of the workflow.
*/
opened_at: string;
/**
* Description
* @description The description of the workflow.
*/
description: string;
/** @description The description of the workflow. */
2024-03-05 11:33:01 +00:00
category: components["schemas"]["WorkflowCategory"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/**
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
* WorkflowRecordOrderBy
* @description The order by options for workflow records
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
WorkflowRecordOrderBy: "created_at" | "updated_at" | "opened_at" | "name";
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** WorkflowWithoutID */
WorkflowWithoutID: {
/**
* Name
* @description The name of the workflow.
*/
name: string;
/**
* Author
* @description The author of the workflow.
*/
author: string;
/**
* Description
* @description The description of the workflow.
*/
description: string;
/**
* Version
* @description The version of the workflow.
*/
version: string;
/**
* Contact
* @description The contact of the workflow.
*/
contact: string;
/**
* Tags
* @description The tags of the workflow.
*/
tags: string;
/**
* Notes
* @description The notes of the workflow.
*/
notes: string;
/**
* Exposedfields
* @description The exposed fields of the workflow.
*/
2024-03-05 11:33:01 +00:00
exposedFields: components["schemas"]["ExposedField"][];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** @description The meta of the workflow. */
2024-03-05 11:33:01 +00:00
meta: components["schemas"]["WorkflowMeta"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Nodes
* @description The nodes of the workflow.
*/
nodes: {
2024-03-05 11:33:01 +00:00
[key: string]: components["schemas"]["JsonValue"];
}[];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/**
* Edges
* @description The edges of the workflow.
*/
edges: {
2024-03-05 11:33:01 +00:00
[key: string]: components["schemas"]["JsonValue"];
}[];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* Zoe (Depth) Processor
* @description Applies Zoe depth processing to image
*/
ZoeDepthImageProcessorInvocation: {
2024-02-07 05:34:40 +00:00
/** @description The board to save the image to */
2024-03-05 11:33:01 +00:00
board?: components["schemas"]["BoardField"] | null;
/** @description Optional metadata to be saved with the image */
2024-03-05 11:33:01 +00:00
metadata?: components["schemas"]["MetadataField"] | null;
/**
2023-08-22 05:15:00 +00:00
* Id
* @description The id of this instance of an invocation. Must be unique among all instances of invocations.
*/
id: string;
/**
2023-08-22 05:15:00 +00:00
* Is Intermediate
* @description Whether or not this is an intermediate invocation.
* @default false
*/
2023-10-18 10:24:07 +00:00
is_intermediate?: boolean;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Use Cache
* @description Whether or not to use the cache
* @default true
*/
use_cache?: boolean;
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/** @description The image to process */
2024-03-05 11:33:01 +00:00
image?: components["schemas"]["ImageField"];
/**
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* type
2023-08-22 05:15:00 +00:00
* @default zoe_depth_image_processor
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @constant
*/
2024-03-05 11:33:01 +00:00
type: "zoe_depth_image_processor";
2023-08-16 01:59:19 +00:00
};
/**
* Classification
* @description The classification of an Invocation.
* - `Stable`: The invocation, including its inputs/outputs and internal logic, is stable. You may build workflows with it, having confidence that they will not break because of a change in this invocation.
* - `Beta`: The invocation is not yet stable, but is planned to be stable in the future. Workflows built around this invocation may break, but we are committed to supporting this invocation long-term.
* - `Prototype`: The invocation is not yet stable and may be removed from the application at any time. Workflows built around this invocation may break, and we are *not* committed to supporting this invocation.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
Classification: "stable" | "beta" | "prototype";
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/**
* FieldKind
* @description The kind of field.
* - `Input`: An input field on a node.
* - `Output`: An output field on a node.
* - `Internal`: A field which is treated as an input, but cannot be used in node definitions. Metadata is
* one example. It is provided to nodes via the WithMetadata class, and we want to reserve the field name
* "metadata" for this on all nodes. `FieldKind` is used to short-circuit the field name validation logic,
* allowing "metadata" for that field.
* - `NodeAttribute`: The field is a node attribute. These are fields which are not inputs or outputs,
* but which are used to store information about the node. For example, the `id` and `type` fields are node
* attributes.
*
* The presence of this in `json_schema_extra["field_kind"]` is used when initializing node schemas on app
* startup, and when generating the OpenAPI schema for the workflow editor.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
FieldKind: "input" | "output" | "internal" | "node_attribute";
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* Input
* @description The type of input a field accepts.
* - `Input.Direct`: The field must have its value provided directly, when the invocation and field are instantiated.
* - `Input.Connection`: The field must have its value provided by a connection.
* - `Input.Any`: The field may have its value provided either directly or by a connection.
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
Input: "connection" | "direct" | "any";
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/**
* InputFieldJSONSchemaExtra
* @description Extra attributes to be added to input fields and their OpenAPI schema. Used during graph execution,
* and by the workflow editor during schema parsing and UI rendering.
*/
InputFieldJSONSchemaExtra: {
2024-03-05 11:33:01 +00:00
input: components["schemas"]["Input"];
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/** Orig Required */
orig_required: boolean;
2024-03-05 11:33:01 +00:00
field_kind: components["schemas"]["FieldKind"];
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/**
* Default
* @default null
*/
default: unknown;
/**
* Orig Default
* @default null
*/
orig_default: unknown;
/**
* Ui Hidden
* @default false
*/
ui_hidden: boolean;
/** @default null */
2024-03-05 11:33:01 +00:00
ui_type: components["schemas"]["UIType"] | null;
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/** @default null */
2024-03-05 11:33:01 +00:00
ui_component: components["schemas"]["UIComponent"] | null;
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/**
* Ui Order
* @default null
*/
ui_order: number | null;
/**
* Ui Choice Labels
* @default null
*/
ui_choice_labels: {
[key: string]: string;
} | null;
};
/**
* OutputFieldJSONSchemaExtra
* @description Extra attributes to be added to input fields and their OpenAPI schema. Used by the workflow editor
* during schema parsing and UI rendering.
*/
OutputFieldJSONSchemaExtra: {
2024-03-05 11:33:01 +00:00
field_kind: components["schemas"]["FieldKind"];
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/** Ui Hidden */
ui_hidden: boolean;
2024-03-05 11:33:01 +00:00
ui_type: components["schemas"]["UIType"] | null;
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
/** Ui Order */
ui_order: number | null;
};
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
/**
* UIComponent
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @description The type of UI component to use for a field, used to override the default components, which are
* inferred from the field type.
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
UIComponent: "none" | "textarea" | "slider";
2023-08-16 01:59:19 +00:00
/**
2023-08-22 05:15:00 +00:00
* UIConfigBase
2023-08-16 01:59:19 +00:00
* @description Provides additional node configuration to the UI.
* This is used internally by the @invocation decorator logic. Do not use this directly.
2023-08-16 01:59:19 +00:00
*/
UIConfigBase: {
/**
2023-08-22 05:15:00 +00:00
* Tags
* @description The node's tags
2023-08-16 01:59:19 +00:00
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
tags: string[] | null;
2023-08-16 01:59:19 +00:00
/**
2023-08-22 05:15:00 +00:00
* Title
* @description The node's display name
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @default null
2023-08-16 01:59:19 +00:00
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
title: string | null;
/**
* Category
* @description The node's category
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
* @default null
*/
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
category: string | null;
/**
* Version
* @description The node's version. Should be a valid semver string e.g. "1.0.0" or "3.8.13".
*/
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
version: string;
/**
* Node Pack
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @description Whether or not this is a custom node
* @default null
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
*/
node_pack: string | null;
/**
* @description The node's classification
* @default stable
*/
2024-03-05 11:33:01 +00:00
classification: components["schemas"]["Classification"];
2023-08-16 01:59:19 +00:00
};
/**
2023-08-22 05:15:00 +00:00
* UIType
feat(ui): add support for custom field types Node authors may now create their own arbitrary/custom field types. Any pydantic model is supported. Two notes: 1. Your field type's class name must be unique. Suggest prefixing fields with something related to the node pack as a kind of namespace. 2. Custom field types function as connection-only fields. For example, if your custom field has string attributes, you will not get a text input for that attribute when you give a node a field with your custom type. This is the same behaviour as other complex fields that don't have custom UIs in the workflow editor - like, say, a string collection. feat(ui): fix tooltips for custom types We need to hold onto the original type of the field so they don't all just show up as "Unknown". fix(ui): fix ts error with custom fields feat(ui): custom field types connection validation In the initial commit, a custom field's original type was added to the *field templates* only as `originalType`. Custom fields' `type` property was `"Custom"`*. This allowed for type safety throughout the UI logic. *Actually, it was `"Unknown"`, but I changed it to custom for clarity. Connection validation logic, however, uses the *field instance* of the node/field. Like the templates, *field instances* with custom types have their `type` set to `"Custom"`, but they didn't have an `originalType` property. As a result, all custom fields could be connected to all other custom fields. To resolve this, we need to add `originalType` to the *field instances*, then switch the validation logic to use this instead of `type`. This ended up needing a bit of fanagling: - If we make `originalType` a required property on field instances, existing workflows will break during connection validation, because they won't have this property. We'd need a new layer of logic to migrate the workflows, adding the new `originalType` property. While this layer is probably needed anyways, typing `originalType` as optional is much simpler. Workflow migration logic can come layer. (Technically, we could remove all references to field types from the workflow files, and let the templates hold all this information. This feels like a significant change and I'm reluctant to do it now.) - Because `originalType` is optional, anywhere we care about the type of a field, we need to use it over `type`. So there are a number of `field.originalType ?? field.type` expressions. This is a bit of a gotcha, we'll need to remember this in the future. - We use `Array.prototype.includes()` often in the workflow editor, e.g. `COLLECTION_TYPES.includes(type)`. In these cases, the const array is of type `FieldType[]`, and `type` is is `FieldType`. Because we now support custom types, the arg `type` is now widened from `FieldType` to `string`. This causes a TS error. This behaviour is somewhat controversial (see https://github.com/microsoft/TypeScript/issues/14520). These expressions are now rewritten as `COLLECTION_TYPES.some((t) => t === type)` to satisfy TS. It's logically equivalent. fix(ui): typo feat(ui): add CustomCollection and CustomPolymorphic field types feat(ui): add validation for CustomCollection & CustomPolymorphic types - Update connection validation for custom types - Use simple string parsing to determine if a field is a collection or polymorphic type. - No longer need to keep a list of collection and polymorphic types. - Added runtime checks in `baseinvocation.py` to ensure no fields are named in such a way that it could mess up the new parsing chore(ui): remove errant console.log fix(ui): rename 'nodes.currentConnectionFieldType' -> 'nodes.connectionStartFieldType' This was confusingly named and kept tripping me up. Renamed to be consistent with the `reactflow` `ConnectionStartParams` type. fix(ui): fix ts error feat(nodes): add runtime check for custom field names "Custom", "CustomCollection" and "CustomPolymorphic" are reserved field names. chore(ui): add TODO for revising field type names wip refactor fieldtype structured wip refactor field types wip refactor types wip refactor types fix node layout refactor field types chore: mypy organisation organisation organisation fix(nodes): fix field orig_required, field_kind and input statuses feat(nodes): remove broken implementation of default_factory on InputField Use of this could break connection validation due to the difference in node schemas required fields and invoke() required args. Removed entirely for now. It wasn't ever actually used by the system, because all graphs always had values provided for fields where default_factory was used. Also, pydantic is smart enough to not reuse the same object when specifying a default value - it clones the object first. So, the common pattern of `default_factory=list` is extraneous. It can just be `default=[]`. fix(nodes): fix InputField name validation workflow validation validation chore: ruff feat(nodes): fix up baseinvocation comments fix(ui): improve typing & logic of buildFieldInputTemplate improved error handling in parseFieldType fix: back compat for deprecated default_factory and UIType feat(nodes): do not show node packs loaded log if none loaded chore(ui): typegen
2023-11-17 00:32:35 +00:00
* @description Type hints for the UI for situations in which the field type is not enough to infer the correct UI type.
*
* - Model Fields
* The most common node-author-facing use will be for model fields. Internally, there is no difference
* between SD-1, SD-2 and SDXL model fields - they all use the class `MainModelField`. To ensure the
* base-model-specific UI is rendered, use e.g. `ui_type=UIType.SDXLMainModelField` to indicate that
* the field is an SDXL main model field.
*
* - Any Field
* We cannot infer the usage of `typing.Any` via schema parsing, so you *must* use `ui_type=UIType.Any` to
* indicate that the field accepts any type. Use with caution. This cannot be used on outputs.
*
* - Scheduler Field
* Special handling in the UI is needed for this field, which otherwise would be parsed as a plain enum field.
*
* - Internal Fields
* Similar to the Any Field, the `collect` and `iterate` nodes make use of `typing.Any`. To facilitate
* handling these types in the client, we use `UIType._Collection` and `UIType._CollectionItem`. These
* should not be used by node authors.
*
* - DEPRECATED Fields
* These types are deprecated and should not be used by node authors. A warning will be logged if one is
* used, and the type will be ignored. They are included here for backwards compatibility.
2023-08-16 01:59:19 +00:00
* @enum {string}
*/
2024-03-05 11:33:01 +00:00
UIType: "SDXLMainModelField" | "SDXLRefinerModelField" | "ONNXModelField" | "VAEModelField" | "LoRAModelField" | "ControlNetModelField" | "IPAdapterModelField" | "SchedulerField" | "AnyField" | "CollectionField" | "CollectionItemField" | "DEPRECATED_Boolean" | "DEPRECATED_Color" | "DEPRECATED_Conditioning" | "DEPRECATED_Control" | "DEPRECATED_Float" | "DEPRECATED_Image" | "DEPRECATED_Integer" | "DEPRECATED_Latents" | "DEPRECATED_String" | "DEPRECATED_BooleanCollection" | "DEPRECATED_ColorCollection" | "DEPRECATED_ConditioningCollection" | "DEPRECATED_ControlCollection" | "DEPRECATED_FloatCollection" | "DEPRECATED_ImageCollection" | "DEPRECATED_IntegerCollection" | "DEPRECATED_LatentsCollection" | "DEPRECATED_StringCollection" | "DEPRECATED_BooleanPolymorphic" | "DEPRECATED_ColorPolymorphic" | "DEPRECATED_ConditioningPolymorphic" | "DEPRECATED_ControlPolymorphic" | "DEPRECATED_FloatPolymorphic" | "DEPRECATED_ImagePolymorphic" | "DEPRECATED_IntegerPolymorphic" | "DEPRECATED_LatentsPolymorphic" | "DEPRECATED_StringPolymorphic" | "DEPRECATED_MainModel" | "DEPRECATED_UNet" | "DEPRECATED_Vae" | "DEPRECATED_CLIP" | "DEPRECATED_Collection" | "DEPRECATED_CollectionItem" | "DEPRECATED_Enum" | "DEPRECATED_WorkflowField" | "DEPRECATED_IsIntermediate" | "DEPRECATED_BoardField" | "DEPRECATED_MetadataItem" | "DEPRECATED_MetadataItemCollection" | "DEPRECATED_MetadataItemPolymorphic" | "DEPRECATED_MetadataDict";
};
responses: never;
parameters: never;
requestBodies: never;
headers: never;
pathItems: never;
};
2024-03-05 11:33:01 +00:00
export type $defs = Record<string, never>;
2023-08-22 05:15:00 +00:00
2024-03-05 11:33:01 +00:00
export type external = Record<string, never>;
export type operations = {
2024-03-05 11:33:01 +00:00
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Parse Dynamicprompts
* @description Creates a batch process
*/
parse_dynamicprompts: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_parse_dynamicprompts"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["DynamicPromptsResponse"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
2024-02-15 11:15:21 +00:00
* List Model Records
* @description Get a list of models.
*/
2024-02-15 11:15:21 +00:00
list_model_records: {
parameters: {
query?: {
/** @description Base models to include */
2024-03-05 11:33:01 +00:00
base_models?: components["schemas"]["BaseModelType"][] | null;
/** @description The type of model to get */
2024-03-05 11:33:01 +00:00
model_type?: components["schemas"]["ModelType"] | null;
2024-02-15 11:15:21 +00:00
/** @description Exact match on the name of the model */
model_name?: string | null;
/** @description Exact match on the format of the model (e.g. 'diffusers') */
2024-03-05 11:33:01 +00:00
model_format?: components["schemas"]["ModelFormat"] | null;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ModelsList"];
2024-02-15 11:15:21 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2024-02-15 11:15:21 +00:00
};
};
};
};
2024-02-22 22:53:33 +00:00
/**
* Get Model Records By Attrs
* @description Gets a model by its attributes. The main use of this route is to provide backwards compatibility with the old
* model manager, which identified models by a combination of name, base and type.
*/
get_model_records_by_attrs: {
parameters: {
query: {
/** @description The name of the model */
name: string;
/** @description The type of the model */
2024-03-05 11:33:01 +00:00
type: components["schemas"]["ModelType"];
2024-02-22 22:53:33 +00:00
/** @description The base model of the model */
2024-03-05 11:33:01 +00:00
base: components["schemas"]["BaseModelType"];
2024-02-22 22:53:33 +00:00
};
};
responses: {
/** @description Successful Response */
200: {
content: {
"application/json": components["schemas"]["MainDiffusersConfig"] | components["schemas"]["MainCheckpointConfig"] | components["schemas"]["VAEDiffusersConfig"] | components["schemas"]["VAECheckpointConfig"] | components["schemas"]["ControlNetDiffusersConfig"] | components["schemas"]["ControlNetCheckpointConfig"] | components["schemas"]["LoRALyCORISConfig"] | components["schemas"]["LoRADiffusersConfig"] | components["schemas"]["TextualInversionFileConfig"] | components["schemas"]["TextualInversionFolderConfig"] | components["schemas"]["IPAdapterConfig"] | components["schemas"]["T2IAdapterConfig"] | components["schemas"]["CLIPVisionDiffusersConfig"];
2024-02-22 22:53:33 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2024-02-22 22:53:33 +00:00
};
};
};
};
2024-02-15 11:15:21 +00:00
/**
* Get Model Record
* @description Get a model record
2024-02-15 11:15:21 +00:00
*/
get_model_record: {
parameters: {
path: {
/** @description Key of the model record to fetch. */
key: string;
};
};
responses: {
/** @description The model configuration was retrieved successfully */
200: {
content: {
"application/json": components["schemas"]["MainDiffusersConfig"] | components["schemas"]["MainCheckpointConfig"] | components["schemas"]["VAEDiffusersConfig"] | components["schemas"]["VAECheckpointConfig"] | components["schemas"]["ControlNetDiffusersConfig"] | components["schemas"]["ControlNetCheckpointConfig"] | components["schemas"]["LoRALyCORISConfig"] | components["schemas"]["LoRADiffusersConfig"] | components["schemas"]["TextualInversionFileConfig"] | components["schemas"]["TextualInversionFolderConfig"] | components["schemas"]["IPAdapterConfig"] | components["schemas"]["T2IAdapterConfig"] | components["schemas"]["CLIPVisionDiffusersConfig"];
};
};
2024-02-15 11:15:21 +00:00
/** @description Bad request */
400: {
content: never;
};
/** @description The model could not be found */
404: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
* Delete Model
2024-02-15 11:15:21 +00:00
* @description Delete model record from database.
*
* The configuration record will be removed. The corresponding weights files will be
* deleted as well if they reside within the InvokeAI "models" directory.
*/
delete_model: {
parameters: {
path: {
2024-02-15 11:15:21 +00:00
/** @description Unique key of model to remove from model registry. */
key: string;
};
};
responses: {
/** @description Model deleted successfully */
2023-08-22 05:15:00 +00:00
204: {
content: never;
};
/** @description Model not found */
2023-08-22 05:15:00 +00:00
404: {
content: never;
};
2023-07-04 05:04:01 +00:00
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2023-07-04 05:04:01 +00:00
};
};
};
};
/**
2024-02-15 11:15:21 +00:00
* Update Model Record
* @description Update a model's config.
2023-07-04 05:04:01 +00:00
*/
2024-02-15 11:15:21 +00:00
update_model_record: {
2023-07-04 05:04:01 +00:00
parameters: {
path: {
2024-02-15 11:15:21 +00:00
/** @description Unique key of model */
key: string;
};
};
requestBody: {
content: {
2024-02-15 11:15:21 +00:00
/**
* @example {
* "path": "/path/to/model",
* "name": "model_name",
* "base": "sd-1",
* "type": "main",
* "format": "checkpoint",
* "config_path": "configs/stable-diffusion/v1-inference.yaml",
2024-02-15 11:15:21 +00:00
* "description": "Model description",
* "variant": "normal"
* }
*/
"application/json": components["schemas"]["ModelRecordChanges"];
2023-07-04 05:04:01 +00:00
};
};
responses: {
/** @description The model was updated successfully */
200: {
2023-07-04 05:04:01 +00:00
content: {
"application/json": components["schemas"]["MainDiffusersConfig"] | components["schemas"]["MainCheckpointConfig"] | components["schemas"]["VAEDiffusersConfig"] | components["schemas"]["VAECheckpointConfig"] | components["schemas"]["ControlNetDiffusersConfig"] | components["schemas"]["ControlNetCheckpointConfig"] | components["schemas"]["LoRALyCORISConfig"] | components["schemas"]["LoRADiffusersConfig"] | components["schemas"]["TextualInversionFileConfig"] | components["schemas"]["TextualInversionFolderConfig"] | components["schemas"]["IPAdapterConfig"] | components["schemas"]["T2IAdapterConfig"] | components["schemas"]["CLIPVisionDiffusersConfig"];
};
};
/** @description Bad request */
2023-08-22 05:15:00 +00:00
400: {
content: never;
};
2023-07-04 05:04:01 +00:00
/** @description The model could not be found */
2023-08-22 05:15:00 +00:00
404: {
content: never;
};
2023-07-17 01:50:35 +00:00
/** @description There is already a model corresponding to the new name */
2023-08-22 05:15:00 +00:00
409: {
content: never;
};
2023-07-15 15:06:57 +00:00
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2023-07-15 15:06:57 +00:00
};
};
};
};
/** Scan For Models */
scan_for_models: {
parameters: {
query?: {
/** @description Directory path to search for models */
scan_path?: string;
};
};
responses: {
/** @description Directory scanned successfully */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["FoundModel"][];
};
};
/** @description Invalid directory path */
400: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
2024-01-22 22:28:24 +00:00
/**
* List Model Installs
* @description Return the list of model install jobs.
*
* Install jobs have a numeric `id`, a `status`, and other fields that provide information on
* the nature of the job and its progress. The `status` is one of:
*
* * "waiting" -- Job is waiting in the queue to run
* * "downloading" -- Model file(s) are downloading
* * "running" -- Model has downloaded and the model probing and registration process is running
* * "completed" -- Installation completed successfully
* * "error" -- An error occurred. Details will be in the "error_type" and "error" fields.
* * "cancelled" -- Job was cancelled before completion.
*
* Once completed, information about the model such as its size, base
* model and type can be retrieved from the `config_out` field. For multi-file models such as diffusers,
* information on individual files can be retrieved from `download_parts`.
*
* See the example and schema below for more information.
2024-01-22 22:28:24 +00:00
*/
list_model_installs: {
2024-01-22 22:28:24 +00:00
responses: {
/** @description Successful Response */
200: {
content: {
"application/json": components["schemas"]["ModelInstallJob"][];
2023-11-16 00:21:28 +00:00
};
};
};
};
/**
* Install Model
2024-02-15 11:15:21 +00:00
* @description Install a model using a string identifier.
*
* `source` can be any of the following.
*
* 1. A path on the local filesystem ('C:\users\fred\model.safetensors')
* 2. A Url pointing to a single downloadable model file
* 3. A HuggingFace repo_id with any of the following formats:
* - model/name
* - model/name:fp16:vae
* - model/name::vae -- use default precision
* - model/name:fp16:path/to/model.safetensors
* - model/name::path/to/model.safetensors
*
* `config` is an optional dict containing model configuration values that will override
* the ones that are probed automatically.
*
* `access_token` is an optional access token for use with Urls that require
* authentication.
*
* Models will be downloaded, probed, configured and installed in a
* series of background threads. The return object has `status` attribute
* that can be used to monitor progress.
*
* See the documentation for `import_model_record` for more information on
* interpreting the job information returned by this route.
*/
install_model: {
2024-02-15 11:15:21 +00:00
parameters: {
query: {
/** @description Model source to install, can be a local path, repo_id, or remote URL */
2024-02-15 11:15:21 +00:00
source: string;
/** @description Whether or not to install a local model in place */
inplace?: boolean | null;
2024-02-15 11:15:21 +00:00
access_token?: string | null;
};
};
requestBody?: {
content: {
/**
* @example {
* "name": "string",
* "description": "string"
2024-02-15 11:15:21 +00:00
* }
*/
2024-03-05 11:33:01 +00:00
"application/json": Record<string, never> | null;
2024-02-15 11:15:21 +00:00
};
};
responses: {
2024-02-15 11:15:21 +00:00
/** @description The model imported successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ModelInstallJob"];
};
};
2024-02-15 11:15:21 +00:00
/** @description There is already a model corresponding to this path or repo_id */
409: {
content: never;
};
/** @description Unrecognized file/folder format */
415: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2024-02-15 11:15:21 +00:00
};
};
/** @description The model appeared to import successfully, but could not be found in the model manager */
424: {
content: never;
};
};
};
/**
* Prune Model Install Jobs
* @description Prune all completed and errored jobs from the install job list.
*/
prune_model_install_jobs: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description All completed and errored jobs have been pruned */
204: {
content: never;
};
/** @description Bad request */
400: {
content: never;
};
};
};
2024-01-22 22:28:24 +00:00
/**
* Get Model Install Job
2024-02-15 11:15:21 +00:00
* @description Return model install job corresponding to the given source. See the documentation for 'List Model Install Jobs'
* for information on the format of the return value.
2024-01-22 22:28:24 +00:00
*/
get_model_install_job: {
parameters: {
path: {
/** @description Model install id */
id: number;
};
};
responses: {
/** @description Success */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ModelInstallJob"];
2024-01-22 22:28:24 +00:00
};
};
/** @description No such job */
404: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2024-01-22 22:28:24 +00:00
};
};
};
};
/**
* Cancel Model Install Job
* @description Cancel the model install job(s) corresponding to the given job ID.
*/
cancel_model_install_job: {
parameters: {
path: {
/** @description Model install job ID */
id: number;
};
};
responses: {
/** @description The job was cancelled successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
2024-01-22 22:28:24 +00:00
};
};
/** @description No such job */
415: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2024-01-22 22:28:24 +00:00
};
};
};
};
/**
* Sync Models To Config
2024-01-22 22:28:24 +00:00
* @description Traverse the models and autoimport directories.
*
* Model files without a corresponding
* record in the database are added. Orphan records without a models file are deleted.
*/
sync_models_to_config: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description Model config record database resynced with files on disk */
204: {
content: never;
};
/** @description Bad request */
400: {
content: never;
};
};
};
2024-02-15 11:15:21 +00:00
/**
* Convert Model
* @description Permanently convert a model into diffusers format, replacing the safetensors version.
* Note that during the conversion process the key and model hash will change.
* The return value is the model configuration for the converted model.
*/
convert_model: {
parameters: {
path: {
/** @description Unique key of the safetensors main model to convert to diffusers format. */
key: string;
};
};
responses: {
/** @description Model converted successfully */
200: {
content: {
"application/json": components["schemas"]["MainDiffusersConfig"] | components["schemas"]["MainCheckpointConfig"] | components["schemas"]["VAEDiffusersConfig"] | components["schemas"]["VAECheckpointConfig"] | components["schemas"]["ControlNetDiffusersConfig"] | components["schemas"]["ControlNetCheckpointConfig"] | components["schemas"]["LoRALyCORISConfig"] | components["schemas"]["LoRADiffusersConfig"] | components["schemas"]["TextualInversionFileConfig"] | components["schemas"]["TextualInversionFolderConfig"] | components["schemas"]["IPAdapterConfig"] | components["schemas"]["T2IAdapterConfig"] | components["schemas"]["CLIPVisionDiffusersConfig"];
2024-02-09 20:14:15 +00:00
};
};
2024-02-15 11:15:21 +00:00
/** @description Bad request */
400: {
content: never;
};
/** @description Model not found */
404: {
content: never;
};
/** @description There is already a model registered at this location */
409: {
content: never;
};
2024-02-09 20:14:15 +00:00
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2024-02-09 20:14:15 +00:00
};
};
};
};
/**
* List Downloads
* @description Get a list of active and inactive jobs.
*/
list_downloads: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["DownloadJob"][];
};
};
};
};
/**
* Prune Downloads
* @description Prune completed and errored jobs.
*/
prune_downloads: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description All completed jobs have been pruned */
204: {
content: never;
};
/** @description Bad request */
400: {
content: never;
};
};
};
/**
* Download
* @description Download the source URL to the file or directory indicted in dest.
*/
download: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_download"];
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["DownloadJob"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
* Get Download Job
* @description Get a download job using its ID.
*/
get_download_job: {
parameters: {
path: {
/** @description ID of the download job to fetch. */
id: number;
};
};
responses: {
/** @description Success */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["DownloadJob"];
};
};
/** @description The requested download JobID could not be found */
404: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
* Cancel Download Job
* @description Cancel a download job using its ID.
*/
cancel_download_job: {
parameters: {
path: {
/** @description ID of the download job to cancel. */
id: number;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description Job has been cancelled */
204: {
content: never;
};
/** @description The requested download JobID could not be found */
404: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
* Cancel All Download Jobs
* @description Cancel all download jobs.
*/
cancel_all_download_jobs: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description Download jobs have been cancelled */
204: {
content: never;
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Upload Image
* @description Uploads an image
*/
upload_image: {
parameters: {
query: {
/** @description The category of the image */
2024-03-05 11:33:01 +00:00
image_category: components["schemas"]["ImageCategory"];
/** @description Whether this is an intermediate image */
is_intermediate: boolean;
2023-07-21 07:45:12 +00:00
/** @description The board to add this image to, if any */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
board_id?: string | null;
/** @description The session ID associated with this upload, if any */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
session_id?: string | null;
/** @description Whether to crop the image */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
crop_visible?: boolean | null;
};
};
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"multipart/form-data": components["schemas"]["Body_upload_image"];
};
};
responses: {
/** @description The image was uploaded successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImageDTO"];
};
};
/** @description Image upload failed */
2023-08-22 05:15:00 +00:00
415: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Get Image Dto
2023-07-12 15:15:09 +00:00
* @description Gets an image's DTO
*/
2023-07-12 15:15:09 +00:00
get_image_dto: {
parameters: {
path: {
2023-07-12 15:15:09 +00:00
/** @description The name of image to get */
image_name: string;
};
};
responses: {
2023-07-12 15:15:09 +00:00
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImageDTO"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Delete Image
* @description Deletes an image
*/
delete_image: {
parameters: {
path: {
/** @description The name of the image to delete */
image_name: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Update Image
* @description Updates an image
*/
update_image: {
parameters: {
path: {
/** @description The name of the image to update */
image_name: string;
};
};
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImageRecordChanges"];
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImageDTO"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
* Get Intermediates Count
* @description Gets the count of intermediate images
*/
get_intermediates_count: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": number;
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Clear Intermediates
* @description Clears all intermediates
*/
clear_intermediates: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": number;
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Get Image Metadata
* @description Gets an image's metadata
*/
get_image_metadata: {
parameters: {
path: {
/** @description The name of image to get */
image_name: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["MetadataField"] | null;
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
/** Get Image Workflow */
get_image_workflow: {
parameters: {
path: {
/** @description The name of image whose workflow to get */
image_name: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["WorkflowWithoutID"] | null;
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
};
};
2023-07-12 15:15:09 +00:00
/**
2023-08-22 05:15:00 +00:00
* Get Image Full
2023-07-12 15:15:09 +00:00
* @description Gets a full-resolution image file
*/
get_image_full: {
parameters: {
path: {
/** @description The name of full-resolution image file to get */
image_name: string;
};
};
responses: {
/** @description Return the full-resolution image */
200: {
content: {
2024-03-05 11:33:01 +00:00
"image/png": unknown;
2023-07-12 15:15:09 +00:00
};
};
/** @description Image not found */
2023-08-22 05:15:00 +00:00
404: {
content: never;
};
2023-07-12 15:15:09 +00:00
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2023-07-12 15:15:09 +00:00
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Get Image Thumbnail
* @description Gets a thumbnail image file
*/
get_image_thumbnail: {
parameters: {
path: {
/** @description The name of thumbnail image file to get */
image_name: string;
};
};
responses: {
/** @description Return the image thumbnail */
200: {
content: {
2024-03-05 11:33:01 +00:00
"image/webp": unknown;
};
};
/** @description Image not found */
2023-08-22 05:15:00 +00:00
404: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Get Image Urls
* @description Gets an image and thumbnail URL
*/
get_image_urls: {
parameters: {
path: {
/** @description The name of the image whose URL to get */
image_name: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImageUrlsDTO"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/**
2023-08-22 05:15:00 +00:00
* List Image Dtos
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Gets a list of image DTOs
*/
list_image_dtos: {
parameters: {
query?: {
/** @description The origin of images to list. */
2024-03-05 11:33:01 +00:00
image_origin?: components["schemas"]["ResourceOrigin"] | null;
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** @description The categories of image to include. */
2024-03-05 11:33:01 +00:00
categories?: components["schemas"]["ImageCategory"][] | null;
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** @description Whether to list intermediate images. */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
is_intermediate?: boolean | null;
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** @description The board id to filter by. Use 'none' to find images without a board. */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
board_id?: string | null;
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/** @description The page offset */
offset?: number;
/** @description The number of images per page */
limit?: number;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["OffsetPaginatedResults_ImageDTO_"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
};
};
/** Delete Images From List */
delete_images_from_list: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_delete_images_from_list"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["DeleteImagesFromListResult"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
};
};
/** Star Images In List */
star_images_in_list: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_star_images_in_list"];
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImagesUpdatedFromListResult"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/** Unstar Images In List */
unstar_images_in_list: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_unstar_images_in_list"];
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImagesUpdatedFromListResult"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/** Download Images From List */
download_images_from_list: {
2024-02-20 11:15:14 +00:00
requestBody?: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_download_images_from_list"];
};
};
responses: {
/** @description Successful Response */
2024-02-20 11:15:14 +00:00
202: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ImagesDownloaded"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
2024-02-20 11:15:14 +00:00
/**
* Get Bulk Download Item
* @description Gets a bulk download zip file
*/
get_bulk_download_item: {
parameters: {
path: {
/** @description The bulk_download_item_name of the bulk download item to get */
bulk_download_item_name: string;
};
};
responses: {
/** @description Return the complete bulk download item */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/zip": unknown;
2024-02-20 11:15:14 +00:00
};
};
/** @description Image not found */
404: {
content: never;
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2024-02-20 11:15:14 +00:00
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* List Boards
* @description Gets a list of boards
*/
list_boards: {
parameters: {
query?: {
/** @description Whether to list all boards */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
all?: boolean | null;
/** @description The page offset */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
offset?: number | null;
/** @description The number of boards per page */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
limit?: number | null;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["OffsetPaginatedResults_BoardDTO_"] | components["schemas"]["BoardDTO"][];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Create Board
* @description Creates a board
*/
create_board: {
parameters: {
query: {
/** @description The name of the board to create */
board_name: string;
};
};
responses: {
/** @description The board was created successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["BoardDTO"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Get Board
* @description Gets a board
*/
get_board: {
parameters: {
path: {
/** @description The id of board to get */
board_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["BoardDTO"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Delete Board
* @description Deletes a board
*/
delete_board: {
parameters: {
query?: {
/** @description Permanently delete all images on the board */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
include_images?: boolean | null;
};
path: {
/** @description The id of board to delete */
board_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["DeleteBoardResult"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Update Board
* @description Updates a board
*/
update_board: {
parameters: {
path: {
/** @description The id of board to update */
board_id: string;
};
};
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["BoardChanges"];
};
};
responses: {
/** @description The board was updated successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["BoardDTO"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* List All Board Image Names
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
* @description Gets a list of images for a board
*/
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
list_all_board_image_names: {
parameters: {
path: {
/** @description The id of the board */
board_id: string;
};
};
responses: {
feat(ui): another go at gallery (#3791) * feat(ui): migrate listImages to RTK query using createEntityAdapter - see comments in `endpoints/images.ts` for explanation of the caching - so far, only manually updating `all` images when new image is generated. no other manual cache updates are implemented, but will be needed. - fixed some weirdness with loading state components (like the spinners in gallery) - added `useThumbnailFallback` for `IAIDndImage`, this displays the tiny webp thumbnail while the full-size images load - comment out some old thunk related stuff in gallerySlice, which is no longer needed * feat(ui): add manual cache updates for board changes (wip) - update RTK Query caches when adding/removing single image to/from board - work more on migrating all image-related operations to RTK Query * update AddImagesToBoardContext so that it works when user uses context menu + modal * handle case where no image is selected * get assets working for main list and boards - dnd only * feat(ui): migrate image uploads to RTK Query - minor refactor of `ImageUploader` and `useImageUploadButton` hooks, simplify some logic - style filesystem upload overlay to match existing UI - replace all old `imageUploaded` thunks with `uploadImage` RTK Query calls, update associated logic including canvas related uploads - simplify `PostUploadAction`s that only need to display user input * feat(ui): remove `receivedPageOfImages` thunks * feat(ui): remove `receivedImageUrls` thunk * feat(ui): finish removing all images thunks stuff now broken: - image usage - delete board images - on first load, no image selected * feat(ui): simplify `updateImage` cache manipulation - we don't actually ever change categories, so we can remove a lot of logic * feat(ui): simplify canvas autosave - instead of using a network request to set the canvas generation as not intermediate, we can just do that in the graph * feat(ui): simplify & handle edge cases in cache updates * feat(db, api): support `board_id='none'` for `get_many` images queries This allows us to get all images that are not on a board. * chore(ui): regen types * feat(ui): add `All Assets`, `No Board` boards Restructure boards: - `all images` is all images - `all assets` is all assets - `no board` is all images/assets without a board set - user boards may have images and assets Update caching logic - much simpler without every board having sub-views of images and assets - update drag and drop operations for all possible interactions * chore(ui): regen types * feat(ui): move download to top of context menu * feat(ui): improve drop overlay styles * fix(ui): fix image not selected on first load - listen for first load of all images board, then select the first image * feat(ui): refactor board deletion api changes: - add route to list all image names for a board. this is required to handle board + image deletion. we need to know every image in the board to determine the image usage across the app. this is fetched only when the delete board and images modal is opened so it's as efficient as it can be. - update the delete board route to respond with a list of deleted `board_images` and `images`, as image names. this is needed to perform accurate clientside state & cache updates after deleting. db changes: - remove unused `board_images` service method to get paginated images dtos for a board. this is now done thru the list images endpoint & images service. needs a small logic change on `images.delete_images_on_board` ui changes: - simplify the delete board modal - no context, just minor prop drilling. this is feasible for boards only because the components that need to trigger and manipulate the modal are very close together in the tree - add cache updates for `deleteBoard` & `deleteBoardAndImages` mutations - the only thing we cannot do directly is on `deleteBoardAndImages`, update the `No Board` board. we'd need to insert image dtos that we may not have loaded. instead, i am just invalidating the tags for that `listImages` cache. so when you `deleteBoardAndImages`, the `No Board` will re-fetch the initial image limit. i think this is more efficient than e.g. fetching all image dtos to insert then inserting them. - handle image usage for `deleteBoardAndImages` - update all (i think/hope) the little bits and pieces in the UI to accomodate these changes * fix(ui): fix board selection logic * feat(ui): add delete board modal loading state * fix(ui): use thumbnails for board cover images * fix(ui): fix race condition with board selection when selecting a board that doesn't have any images loaded, we need to wait until the images haveloaded before selecting the first image. this logic is debounced to ~1000ms. * feat(ui): name 'No Board' correctly, change icon * fix(ui): do not cache listAllImageNames query if we cache it, we can end up with stale image usage during deletion. we could of course manually update the cache as we are doing elsewhere. but because this is a relatively infrequent network request, i'd like to trade increased cache mgmt complexity here for increased resource usage. * feat(ui): reduce drag preview opacity, remove border * fix(ui): fix incorrect queryArg used in `deleteImage` and `updateImage` cache updates * fix(ui): fix doubled open in new tab * fix(ui): fix new generations not getting added to 'No Board' * fix(ui): fix board id not changing on new image when autosave enabled * fix(ui): context menu when selection is 0 need to revise how context menu is triggered later, when we approach multi select * fix(ui): fix deleting does not update counts for all images and all assets * fix(ui): fix all assets board name in boards list collapse button * fix(ui): ensure we never go under 0 for total board count * fix(ui): fix text overflow on board names --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-07-19 16:06:38 +00:00
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": string[];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Add Image To Board
* @description Creates a board_image
*/
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
add_image_to_board: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_add_image_to_board"];
};
};
responses: {
/** @description The image was added to a board successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Remove Image From Board
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Removes an image from its board, if it had one
*/
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
remove_image_from_board: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_remove_image_from_board"];
};
};
responses: {
/** @description The image was removed from the board successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
/**
2023-08-22 05:15:00 +00:00
* Add Images To Board
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Adds a list of images to a board
*/
add_images_to_board: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_add_images_to_board"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
responses: {
/** @description Images were added to board successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["AddImagesToBoardResult"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
};
};
/**
2023-08-22 05:15:00 +00:00
* Remove Images From Board
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
* @description Removes a list of images from their board, if they had one
*/
remove_images_from_board: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_remove_images_from_board"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
responses: {
/** @description Images were removed from board successfully */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["RemoveImagesFromBoardResult"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: add multi-select to gallery multi-select actions include: - drag to board to move all to that board - right click to add all to board or delete all backend changes: - add routes for changing board for list of image names, deleting list of images - change image-specific routes to `images/i/{image_name}` to not clobber other routes (like `images/upload`, `images/delete`) - subclass pydantic `BaseModel` as `BaseModelExcludeNull`, which excludes null values when calling `dict()` on the model. this fixes inconsistent types related to JSON parsing null values into `null` instead of `undefined` - remove `board_id` from `remove_image_from_board` frontend changes: - multi-selection stuff uses `ImageDTO[]` as payloads, for dnd and other mutations. this gives us access to image `board_id`s when hitting routes, and enables efficient cache updates. - consolidate change board and delete image modals to handle single and multiples - board totals are now re-fetched on mutation and not kept in sync manually - was way too tedious to do this - fixed warning about nested `<p>` elements - closes #4088 , need to handle case when `autoAddBoardId` is `"none"` - add option to show gallery image delete button on every gallery image frontend refactors/organisation: - make typegen script js instead of ts - enable `noUncheckedIndexedAccess` to help avoid bugs when indexing into arrays, many small changes needed to satisfy TS after this - move all image-related endpoints into `endpoints/images.ts`, its a big file now, but this fixes a number of circular dependency issues that were otherwise felt impossible to resolve
2023-07-31 08:16:52 +00:00
};
};
};
};
2023-07-08 09:31:17 +00:00
/** Get Version */
app_version: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["AppVersion"];
};
};
2023-07-12 15:56:40 +00:00
};
};
2023-12-06 12:33:54 +00:00
/** Get App Deps */
get_app_deps: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["AppDependencyVersions"];
2023-12-06 12:33:54 +00:00
};
};
};
};
2023-07-12 15:56:40 +00:00
/** Get Config */
get_config: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["AppConfig"];
};
};
};
};
2023-07-20 15:45:54 +00:00
/**
2023-08-22 05:15:00 +00:00
* Get Log Level
2023-07-20 15:45:54 +00:00
* @description Returns the log level
*/
get_log_level: {
2023-07-08 09:31:17 +00:00
responses: {
2023-07-20 15:45:54 +00:00
/** @description The operation was successful */
2023-07-08 09:31:17 +00:00
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["LogLevel"];
2023-07-12 15:56:40 +00:00
};
};
};
};
2023-07-20 15:45:54 +00:00
/**
2023-08-22 05:15:00 +00:00
* Set Log Level
2023-07-20 15:45:54 +00:00
* @description Sets the log verbosity level
*/
set_log_level: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["LogLevel"];
2023-07-20 15:45:54 +00:00
};
};
2023-07-12 15:56:40 +00:00
responses: {
2023-07-20 15:45:54 +00:00
/** @description The operation was successful */
2023-07-12 15:56:40 +00:00
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["LogLevel"];
2023-07-20 15:45:54 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
2023-07-08 09:31:17 +00:00
};
};
};
};
/**
* Clear Invocation Cache
* @description Clears the invocation cache
*/
clear_invocation_cache: {
responses: {
/** @description The operation was successful */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
};
};
/**
* Enable Invocation Cache
* @description Clears the invocation cache
*/
enable_invocation_cache: {
responses: {
/** @description The operation was successful */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
};
};
/**
* Disable Invocation Cache
* @description Clears the invocation cache
*/
disable_invocation_cache: {
responses: {
/** @description The operation was successful */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
};
};
};
};
/**
* Get Invocation Cache Status
* @description Clears the invocation cache
*/
get_invocation_cache_status: {
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["InvocationCacheStatus"];
};
};
};
};
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/**
* Enqueue Batch
* @description Processes a batch and enqueues the output graphs for execution.
*/
enqueue_batch: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_enqueue_batch"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["EnqueueBatchResult"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Created */
201: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["EnqueueBatchResult"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* List Queue Items
* @description Gets all queue items (without graphs)
*/
list_queue_items: {
parameters: {
query?: {
/** @description The number of items to fetch */
limit?: number;
/** @description The status of items to fetch */
2024-03-05 11:33:01 +00:00
status?: ("pending" | "in_progress" | "completed" | "failed" | "canceled") | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** @description The pagination cursor */
feat(api): chore: pydantic & fastapi upgrade Upgrade pydantic and fastapi to latest. - pydantic~=2.4.2 - fastapi~=103.2 - fastapi-events~=0.9.1 **Big Changes** There are a number of logic changes needed to support pydantic v2. Most changes are very simple, like using the new methods to serialized and deserialize models, but there are a few more complex changes. **Invocations** The biggest change relates to invocation creation, instantiation and validation. Because pydantic v2 moves all validation logic into the rust pydantic-core, we may no longer directly stick our fingers into the validation pie. Previously, we (ab)used models and fields to allow invocation fields to be optional at instantiation, but required when `invoke()` is called. We directly manipulated the fields and invocation models when calling `invoke()`. With pydantic v2, this is much more involved. Changes to the python wrapper do not propagate down to the rust validation logic - you have to rebuild the model. This causes problem with concurrent access to the invocation classes and is not a free operation. This logic has been totally refactored and we do not need to change the model any more. The details are in `baseinvocation.py`, in the `InputField` function and `BaseInvocation.invoke_internal()` method. In the end, this implementation is cleaner. **Invocation Fields** In pydantic v2, you can no longer directly add or remove fields from a model. Previously, we did this to add the `type` field to invocations. **Invocation Decorators** With pydantic v2, we instead use the imperative `create_model()` API to create a new model with the additional field. This is done in `baseinvocation.py` in the `invocation()` wrapper. A similar technique is used for `invocation_output()`. **Minor Changes** There are a number of minor changes around the pydantic v2 models API. **Protected `model_` Namespace** All models' pydantic-provided methods and attributes are prefixed with `model_` and this is considered a protected namespace. This causes some conflict, because "model" means something to us, and we have a ton of pydantic models with attributes starting with "model_". Forunately, there are no direct conflicts. However, in any pydantic model where we define an attribute or method that starts with "model_", we must tell set the protected namespaces to an empty tuple. ```py class IPAdapterModelField(BaseModel): model_name: str = Field(description="Name of the IP-Adapter model") base_model: BaseModelType = Field(description="Base model") model_config = ConfigDict(protected_namespaces=()) ``` **Model Serialization** Pydantic models no longer have `Model.dict()` or `Model.json()`. Instead, we use `Model.model_dump()` or `Model.model_dump_json()`. **Model Deserialization** Pydantic models no longer have `Model.parse_obj()` or `Model.parse_raw()`, and there are no `parse_raw_as()` or `parse_obj_as()` functions. Instead, you need to create a `TypeAdapter` object to parse python objects or JSON into a model. ```py adapter_graph = TypeAdapter(Graph) deserialized_graph_from_json = adapter_graph.validate_json(graph_json) deserialized_graph_from_dict = adapter_graph.validate_python(graph_dict) ``` **Field Customisation** Pydantic `Field`s no longer accept arbitrary args. Now, you must put all additional arbitrary args in a `json_schema_extra` arg on the field. **Schema Customisation** FastAPI and pydantic schema generation now follows the OpenAPI version 3.1 spec. This necessitates two changes: - Our schema customization logic has been revised - Schema parsing to build node templates has been revised The specific aren't important, but this does present additional surface area for bugs. **Performance Improvements** Pydantic v2 is a full rewrite with a rust backend. This offers a substantial performance improvement (pydantic claims 5x to 50x depending on the task). We'll notice this the most during serialization and deserialization of sessions/graphs, which happens very very often - a couple times per node. I haven't done any benchmarks, but anecdotally, graph execution is much faster. Also, very larges graphs - like with massive iterators - are much, much faster.
2023-09-24 08:11:07 +00:00
cursor?: number | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
/** @description The pagination cursor priority */
priority?: number;
};
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["CursorPaginatedResults_SessionQueueItemDTO_"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Resume
* @description Resumes session processor
*/
resume: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["SessionProcessorStatus"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Pause
* @description Pauses session processor
*/
pause: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["SessionProcessorStatus"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Cancel By Batch Ids
* @description Immediately cancels all queue items from the given batch ids
*/
cancel_by_batch_ids: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_cancel_by_batch_ids"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["CancelByBatchIDsResult"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Clear
* @description Clears the queue entirely, immediately canceling the currently-executing session
*/
clear: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["ClearResult"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Prune
* @description Prunes all completed or errored queue items
*/
prune: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["PruneResult"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Get Current Queue Item
* @description Gets the currently execution queue item
*/
get_current_queue_item: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["SessionQueueItem"] | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Get Next Queue Item
* @description Gets the next queue item, without executing it
*/
get_next_queue_item: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["SessionQueueItem"] | null;
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Get Queue Status
* @description Gets the status of the session queue
*/
get_queue_status: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["SessionQueueAndProcessorStatus"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Get Batch Status
* @description Gets the status of the session queue
*/
get_batch_status: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
/** @description The batch to get the status of */
batch_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["BatchStatus"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Get Queue Item
* @description Gets a queue item
*/
get_queue_item: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
/** @description The queue item to get */
item_id: number;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["SessionQueueItem"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Cancel Queue Item
* @description Deletes a queue item
*/
cancel_queue_item: {
parameters: {
path: {
/** @description The queue id to perform this operation on */
queue_id: string;
/** @description The queue item to cancel */
item_id: number;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["SessionQueueItem"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: queued generation (#4502) * fix(config): fix typing issues in `config/` `config/invokeai_config.py`: - use `Optional` for things that are optional - fix typing of `ram_cache_size()` and `vram_cache_size()` - remove unused and incorrectly typed method `autoconvert_path` - fix types and logic for `parse_args()`, in which `InvokeAIAppConfig.initconf` *must* be a `DictConfig`, but function would allow it to be set as a `ListConfig`, which presumably would cause issues elsewhere `config/base.py`: - use `cls` for first arg of class methods - use `Optional` for things that are optional - fix minor type issue related to setting of `env_prefix` - remove unused `add_subparser()` method, which calls `add_parser()` on an `ArgumentParser` (method only available on the `_SubParsersAction` object, which is returned from ArgumentParser.add_subparsers()`) * feat: queued generation and batches Due to a very messy branch with broad addition of `isort` on `main` alongside it, some git surgery was needed to get an agreeable git history. This commit represents all of the work on queued generation. See PR for notes. * chore: flake8, isort, black * fix(nodes): fix incorrect service stop() method * fix(nodes): improve names of a few variables * fix(tests): fix up tests after changes to batches/queue * feat(tests): add unit tests for session queue helper functions * feat(ui): dynamic prompts is always enabled * feat(queue): add queue_status_changed event * feat(ui): wip queue graphs * feat(nodes): move cleanup til after invoker startup * feat(nodes): add cancel_by_batch_ids * feat(ui): wip batch graphs & UI * fix(nodes): remove `Batch.batch_id` from required * fix(ui): cleanup and use fixedCacheKey for all mutations * fix(ui): remove orphaned nodes from canvas graphs * fix(nodes): fix cancel_by_batch_ids result count * fix(ui): only show cancel batch tooltip when batches were canceled * chore: isort * fix(api): return `[""]` when dynamic prompts generates no prompts Just a simple fallback so we always have a prompt. * feat(ui): dynamicPrompts.combinatorial is always on There seems to be little purpose in using the combinatorial generation for dynamic prompts. I've disabled it by hiding it from the UI and defaulting combinatorial to true. If we want to enable it again in the future it's straightforward to do so. * feat: add queue_id & support logic * feat(ui): fix upscale button It prepends the upscale operation to queue * feat(nodes): return queue item when enqueuing a single graph This facilitates one-off graph async workflows in the client. * feat(ui): move controlnet autoprocess to queue * fix(ui): fix non-serializable DOMRect in redux state * feat(ui): QueueTable performance tweaks * feat(ui): update queue list Queue items expand to show the full queue item. Just as JSON for now. * wip threaded session_processor * feat(nodes,ui): fully migrate queue to session_processor * feat(nodes,ui): add processor events * feat(ui): ui tweaks * feat(nodes,ui): consolidate events, reduce network requests * feat(ui): cleanup & abstract queue hooks * feat(nodes): optimize batch permutation Use a generator to do only as much work as is needed. Previously, though we only ended up creating exactly as many queue items as was needed, there was still some intermediary work that calculated *all* permutations. When that number was very high, the system had a very hard time and used a lot of memory. The logic has been refactored to use a generator. Additionally, the batch validators are optimized to return early and use less memory. * feat(ui): add seed behaviour parameter This dynamic prompts parameter allows the seed to be randomized per prompt or per iteration: - Per iteration: Use the same seed for all prompts in a single dynamic prompt expansion - Per prompt: Use a different seed for every single prompt "Per iteration" is appropriate for exploring a the latents space with a stable starting noise, while "Per prompt" provides more variation. * fix(ui): remove extraneous random seed nodes from linear graphs * fix(ui): fix controlnet autoprocess not working when queue is running * feat(queue): add timestamps to queue status updates Also show execution time in queue list * feat(queue): change all execution-related events to use the `queue_id` as the room, also include `queue_item_id` in InvocationQueueItem This allows for much simpler handling of queue items. * feat(api): deprecate sessions router * chore(backend): tidy logging in `dependencies.py` * fix(backend): respect `use_memory_db` * feat(backend): add `config.log_sql` (enables sql trace logging) * feat: add invocation cache Supersedes #4574 The invocation cache provides simple node memoization functionality. Nodes that use the cache are memoized and not re-executed if their inputs haven't changed. Instead, the stored output is returned. ## Results This feature provides anywhere some significant to massive performance improvement. The improvement is most marked on large batches of generations where you only change a couple things (e.g. different seed or prompt for each iteration) and low-VRAM systems, where skipping an extraneous model load is a big deal. ## Overview A new `invocation_cache` service is added to handle the caching. There's not much to it. All nodes now inherit a boolean `use_cache` field from `BaseInvocation`. This is a node field and not a class attribute, because specific instances of nodes may want to opt in or out of caching. The recently-added `invoke_internal()` method on `BaseInvocation` is used as an entrypoint for the cache logic. To create a cache key, the invocation is first serialized using pydantic's provided `json()` method, skipping the unique `id` field. Then python's very fast builtin `hash()` is used to create an integer key. All implementations of `InvocationCacheBase` must provide a class method `create_key()` which accepts an invocation and outputs a string or integer key. ## In-Memory Implementation An in-memory implementation is provided. In this implementation, the node outputs are stored in memory as python classes. The in-memory cache does not persist application restarts. Max node cache size is added as `node_cache_size` under the `Generation` config category. It defaults to 512 - this number is up for discussion, but given that these are relatively lightweight pydantic models, I think it's safe to up this even higher. Note that the cache isn't storing the big stuff - tensors and images are store on disk, and outputs include only references to them. ## Node Definition The default for all nodes is to use the cache. The `@invocation` decorator now accepts an optional `use_cache: bool` argument to override the default of `True`. Non-deterministic nodes, however, should set this to `False`. Currently, all random-stuff nodes, including `dynamic_prompt`, are set to `False`. The field name `use_cache` is now effectively a reserved field name and possibly a breaking change if any community nodes use this as a field name. In hindsight, all our reserved field names should have been prefixed with underscores or something. ## One Gotcha Leaf nodes probably want to opt out of the cache, because if they are not cached, their outputs are not saved again. If you run the same graph multiple times, you only end up with a single image output, because the image storage side-effects are in the `invoke()` method, which is bypassed if we have a cache hit. ## Linear UI The linear graphs _almost_ just work, but due to the gotcha, we need to be careful about the final image-outputting node. To resolve this, a `SaveImageInvocation` node is added and used in the linear graphs. This node is similar to `ImagePrimitive`, except it saves a copy of its input image, and has `use_cache` set to `False` by default. This is now the leaf node in all linear graphs, and is the only node in those graphs with `use_cache == False` _and_ the only node with `is_intermedate == False`. ## Workflow Editor All nodes now have a footer with a new `Use Cache [ ]` checkbox. It defaults to the value set by the invocation in its python definition, but can be changed by the user. The workflow/node validation logic has been updated to migrate old workflows to use the new default values for `use_cache`. Users may still want to review the settings that have been chosen. In the event of catastrophic failure when running this migration, the default value of `True` is applied, as this is correct for most nodes. Users should consider saving their workflows after loading them in and having them updated. ## Future Enhancements - Callback A future enhancement would be to provide a callback to the `use_cache` flag that would be run as the node is executed to determine, based on its own internal state, if the cache should be used or not. This would be useful for `DynamicPromptInvocation`, where the deterministic behaviour is determined by the `combinatorial: bool` field. ## Future Enhancements - Persisted Cache Similar to how the latents storage is backed by disk, the invocation cache could be persisted to the database or disk. We'd need to be very careful about deserializing outputs, but it's perhaps worth exploring in the future. * fix(ui): fix queue list item width * feat(nodes): do not send the whole node on every generator progress * feat(ui): strip out old logic related to sessions Things like `isProcessing` are no longer relevant with queue. Removed them all & updated everything be appropriate for queue. May be a few little quirks I've missed... * feat(ui): fix up param collapse labels * feat(ui): click queue count to go to queue tab * tidy(queue): update comment, query format * feat(ui): fix progress bar when canceling * fix(ui): fix circular dependency * feat(nodes): bail on node caching logic if `node_cache_size == 0` * feat(nodes): handle KeyError on node cache pop * feat(nodes): bypass cache codepath if caches is disabled more better no do thing * fix(ui): reset api cache on connect/disconnect * feat(ui): prevent enqueue when no prompts generated * feat(ui): add queue controls to workflow editor * feat(ui): update floating buttons & other incidental UI tweaks * fix(ui): fix missing/incorrect translation keys * fix(tests): add config service to mock invocation services invoking needs access to `node_cache_size` to occur * optionally remove pause/resume buttons from queue UI * option to disable prepending * chore(ui): remove unused file * feat(queue): remove `order_id` entirely, `item_id` is now an autoinc pk --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-09-20 05:09:24 +00:00
};
};
};
};
/**
* Get Workflow
* @description Gets a workflow
*/
get_workflow: {
parameters: {
path: {
/** @description The workflow to get */
workflow_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["WorkflowRecordDTO"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
};
};
/**
* Delete Workflow
* @description Deletes a workflow
*/
delete_workflow: {
parameters: {
path: {
/** @description The workflow to delete */
workflow_id: string;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": unknown;
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
};
};
/**
* Update Workflow
* @description Updates a workflow
*/
update_workflow: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_update_workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["WorkflowRecordDTO"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
};
};
/**
* List Workflows
* @description Gets a page of workflows
*/
list_workflows: {
parameters: {
query?: {
/** @description The page to get */
page?: number;
/** @description The number of workflows per page */
per_page?: number;
2023-12-06 12:33:54 +00:00
/** @description The attribute to order by */
2024-03-05 11:33:01 +00:00
order_by?: components["schemas"]["WorkflowRecordOrderBy"];
2023-12-06 12:33:54 +00:00
/** @description The direction to order by */
2024-03-05 11:33:01 +00:00
direction?: components["schemas"]["SQLiteDirection"];
2023-12-06 12:33:54 +00:00
/** @description The category of workflow to get */
2024-03-05 11:33:01 +00:00
category?: components["schemas"]["WorkflowCategory"];
2023-12-06 12:33:54 +00:00
/** @description The text to query by (matches name and description) */
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
query?: string | null;
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["PaginatedResults_WorkflowRecordListItemDTO_"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
};
};
/**
* Create Workflow
* @description Creates a workflow
*/
create_workflow: {
requestBody: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["Body_create_workflow"];
feat: workflow library (#5148) * chore: bump pydantic to 2.5.2 This release fixes pydantic/pydantic#8175 and allows us to use `JsonValue` * fix(ui): exclude public/en.json from prettier config * fix(workflow_records): fix SQLite workflow insertion to ignore duplicates * feat(backend): update workflows handling Update workflows handling for Workflow Library. **Updated Workflow Storage** "Embedded Workflows" are workflows associated with images, and are now only stored in the image files. "Library Workflows" are not associated with images, and are stored only in DB. This works out nicely. We have always saved workflows to files, but recently began saving them to the DB in addition to in image files. When that happened, we stopped reading workflows from files, so all the workflows that only existed in images were inaccessible. With this change, access to those workflows is restored, and no workflows are lost. **Updated Workflow Handling in Nodes** Prior to this change, workflows were embedded in images by passing the whole workflow JSON to a special workflow field on a node. In the node's `invoke()` function, the node was able to access this workflow and save it with the image. This (inaccurately) models workflows as a property of an image and is rather awkward technically. A workflow is now a property of a batch/session queue item. It is available in the InvocationContext and therefore available to all nodes during `invoke()`. **Database Migrations** Added a `SQLiteMigrator` class to handle database migrations. Migrations were needed to accomodate the DB-related changes in this PR. See the code for details. The `images`, `workflows` and `session_queue` tables required migrations for this PR, and are using the new migrator. Other tables/services are still creating tables themselves. A followup PR will adapt them to use the migrator. **Other/Support Changes** - Add a `has_workflow` column to `images` table to indicate that the image has an embedded workflow. - Add handling for retrieving the workflow from an image in python. The image file must be fetched, the workflow extracted, and then sent to client, avoiding needing the browser to parse the image file. With the `has_workflow` column, the UI knows if there is a workflow to be fetched, and only fetches when the user requests to load the workflow. - Add route to get the workflow from an image - Add CRUD service/routes for the library workflows - `workflow_images` table and services removed (no longer needed now that embedded workflows are not in the DB) * feat(ui): updated workflow handling (WIP) Clientside updates for the backend workflow changes. Includes roughed-out workflow library UI. * feat: revert SQLiteMigrator class Will pursue this in a separate PR. * feat(nodes): do not overwrite custom node module names Use a different, simpler method to detect if a node is custom. * feat(nodes): restore WithWorkflow as no-op class This class is deprecated and no longer needed. Set its workflow attr value to None (meaning it is now a no-op), and issue a warning when an invocation subclasses it. * fix(nodes): fix get_workflow from queue item dict func * feat(backend): add WorkflowRecordListItemDTO This is the id, name, description, created at and updated at workflow columns/attrs. Used to display lists of workflowsl * chore(ui): typegen * feat(ui): add workflow loading, deleting to workflow library UI * feat(ui): workflow library pagination button styles * wip * feat: workflow library WIP - Save to library - Duplicate - Filter/sort - UI/queries * feat: workflow library - system graphs - wip * feat(backend): sync system workflows to db * fix: merge conflicts * feat: simplify default workflows - Rename "system" -> "default" - Simplify syncing logic - Update UI to match * feat(workflows): update default workflows - Update TextToImage_SD15 - Add TextToImage_SDXL - Add README * feat(ui): refine workflow list UI * fix(workflow_records): typo * fix(tests): fix tests * feat(ui): clean up workflow library hooks * fix(db): fix mis-ordered db cleanup step It was happening before pruning queue items - should happen afterwards, else you have to restart the app again to free disk space made available by the pruning. * feat(ui): tweak reset workflow editor translations * feat(ui): split out workflow redux state The `nodes` slice is a rather complicated slice. Removing `workflow` makes it a bit more reasonable. Also helps to flatten state out a bit. * docs: update default workflows README * fix: tidy up unused files, unrelated changes * fix(backend): revert unrelated service organisational changes * feat(backend): workflow_records.get_many arg "filter_text" -> "query" * feat(ui): use custom hook in current image buttons Already in use elsewhere, forgot to use it here. * fix(ui): remove commented out property * fix(ui): fix workflow loading - Different handling for loading from library vs external - Fix bug where only nodes and edges loaded * fix(ui): fix save/save-as workflow naming * fix(ui): fix circular dependency * fix(db): fix bug with releasing without lock in db.clean() * fix(db): remove extraneous lock * chore: bump ruff * fix(workflow_records): default `category` to `WorkflowCategory.User` This allows old workflows to validate when reading them from the db or image files. * hide workflow library buttons if feature is disabled --------- Co-authored-by: Mary Hipp <maryhipp@Marys-MacBook-Air.local>
2023-12-08 22:48:38 +00:00
};
};
responses: {
/** @description Successful Response */
200: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["WorkflowRecordDTO"];
};
};
/** @description Validation Error */
422: {
content: {
2024-03-05 11:33:01 +00:00
"application/json": components["schemas"]["HTTPValidationError"];
};
};
};
};
};